Chain-of-Thought

Instead of asking AI to just spit out an answer, getting it to "work through the steps" before answering turns out to sharply improve accuracy. This became famous through a 2022 Google study, and the phrase "Let's think step by step" was passed around like a magic spell.

It's the same logic as why doing math on paper is more accurate than doing it in your head. This discovery evolved further into reasoning models that have "thinking time" built right in — when an article mentions a model's "thought process" or "thinking," this is the lineage it comes from.

1. Try asking this: "I have 5 apples, ate 2, bought 3 more, then gave away half. How many are left? Answer in one word only."
2. Then try: "Solve the same problem step by step, showing your calculations."
3. Notice how writing out the process reduces mistakes — that's chain-of-thought. Modern reasoning models are trained to do this "writing out the process" internally, even without being asked.