How to design a good prompt? Chain of Thought Prompting Elicits Reasoning in Large Language Models Paper (Must read!), First proposes the concept of CoT 2 class of tasks: system-1 tasks (e.g., sentiment analysis, topic classification), system-2 tasks (e.g., logical, mathematical, commonsense reasoning) System-2 tasks using standard prompting have flat scaling curves (the largest model did not achieve high performance)