In a new paper, researchers argue that emergent properties of LLM-powered AIs — widely seen as a step toward AGI — may just be "mirages."