Skip to slide
07 / 53

CH 01 · Chain of Thought, Teaching Models to Think Out Loud · 02/08

Read as one page

The problem: answering too fast

Imagine asking a person this question and demanding an instant, one-word answer:

A cafe had 23 apples. They used 20 to make pies and then bought 6 more. How many apples do they have now?

If forced to blurt out a number immediately, even a smart person might slip. But given a moment to think, "23 minus 20 is 3, plus 6 is 9," they get it right easily. Language models have the same issue, and for a deep reason. A model produces its answer one token at a time, with a fixed, small amount of computation per token. If you ask it to jump straight to the final number, it has almost no room to work out the intermediate steps. It is being forced to answer too fast.

Finished this idea?

Your learning progress stays in this browser.