Jargon:Chain of thought
1 min read
Say this at the party
“Chain of thought is when an AI works through a problem step by step before answering, like showing your working in a maths test.”
In plain English
If you ask an AI a tricky question and make it answer straight away, it often guesses. If it writes out the steps first, it gets far more right.
That written-out reasoning is called a chain of thought. Each step gives the next one something to build on.
It started as a prompting trick. Now many AIs, often called reasoning models, are trained to do it on their own before they reply.
Where it came from
The phrase comes from a 2022 Google paper by Jason Wei and colleagues, which showed that giving a model a few worked examples made it much better at maths and logic puzzles. Later that year, another team found that simply adding "Let's think step by step" helped too.
See it in action
Maya asks: "I have three pizzas cut into eight slices. Five of us eat three slices each. How many are left?" With chain of thought, the AI works out 24 slices, minus 15 eaten, and says 9.
What people get wrong
- The steps an AI shows aren't always how it really got the answer. Treat them as a helpful explanation, not proof.
- More thinking isn't always better. For simple questions it just costs more time and tokens.
Go deeper
- PaperWei et al., Google, 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language ModelsThe paper that named chain of thought and showed how much it helps.
- PaperKojima et al., 2022Large Language Models are Zero-Shot ReasonersThe "Let's think step by step" paper. One sentence, big gains.
- PostGoogle ResearchLanguage Models Perform Reasoning via Chain of ThoughtGoogle's own easy-reading write-up of the idea, with examples.
Related jargon
Drafted with AI. The AI Council (Claude, ChatGPT and Gemini) gave counsel. Phin signed it off.