Chapter 13 · Reasoning models2024
Thinking tokens and extended thinking
Reasoning models write private working-out tokens before the final answer.
Chapter 13 · Reasoning models2024
Reasoning models write private working-out tokens before the final answer.
A bat and a ball cost ₹110 together. The bat costs ₹100 more than the ball. How much is the ball?
A reasoning model writes working-out tokens first. They are often hidden from you, but the model reads them, and they take time.
You met tokens early on. Everything a model reads or writes is counted in tokens.
Pick a question, then switch between a quick answer and a think-first answer. Watch the token bar and whether the answer is right.
A bat and a ball cost ₹110 together. The bat costs ₹100 more than the ball. How much is the ball?
The quick answer sounds right but fails the check.
The replies are scripted for this lesson. Token counts are rough (about four characters each). Real models can also get a quick answer right, or a long think wrong.
Hidden thinking tokens are…
Which question gains least from thinking first?
Why can thinking first fix a trap question?