AI Anatomy Class · Part 10 · The world after ChatGPT

87. Think, then answer

Q. When you give an AI a hard problem, it shows 'Thinking…'. What is it doing all that time?

1. Try this: a bat and a ball cost $1.10, and the bat costs $1.00 more than the ball. What does the ball cost? 10¢ jumps to mind.
2. Give this to the model with 'Answer only!'. It has to answer in the very first piece, so the plausible '10¢' gets picked easily.
3. Let's check. If the ball is 10¢, the bat is $1.10, and together $1.20. Wrong! Each piece gets only one run of the model (chapter 2).
4. Now start with 'Think step by step.' If the ball is x cents, the bat is x + 100, so x + (x + 100) = 110.
5. Then 2x = 10, so x = 5. The ball costs 5¢! Rereading each line it wrote, it split the hard math into small steps.
6. Having it write the working first is called chain of thought. A 2022 study found it helped bigger models more.
7. Reasoning models are trained for this. One known way: reward long working that ends in the right answer (chapter 76).
8. But writing many pieces takes longer, and the working itself can be wrong. Check important answers yourself.

Chapter summary

• Answering at once means picking the answer in the first piece, so a plausible wrong answer comes out easily.
• Writing the working first lets the model reread its lines and split the math into steps (chain of thought).
• Reasoning models are trained to think at length before answering. But they're slower, and the working needs checking too.