What is Reasoning?
Reasoning models and chain of thought, told as a story about a parrot that gets time to mutter to itself before it answers.
The parrot answers a math problem right away. Wrong.It did the multiply and skipped the minus. This happens a lot when the answer comes out first.
The parrot adds beans one by one. If the answer bean goes first, there is no room to work.Earlier beans decide later ones. A step not written as beans is a step that never happened — there is no doing it in its head.
Reasoning is giving the parrot time to mutter to itself before it answers.It writes the middle steps down as beans, and picks the next bean while looking at them. So the next bean gets better.
Answering at once takes 3 beans; muttering takes 11. Slower, but right.Thinking beans are still beans — they cost money and time. So you leave it off for easy questions and turn it on for hard ones. How many thinking beans to allow is the thinking budget.
There is no promise the muttering is right. Plausible thoughts are still just continuing.A wrong first step makes the checking step wrong too — the plausible parrot can be plausible in its thoughts as well. And every thinking bean makes the answer slower and the bean bill bigger.
Reasoning = giving the parrot time to mutter to itself before it answers. Writing the middle steps as beans makes the next bean better. Thinking beans cost money and time, so use them only for hard problems — and the answer is still a draft.
Having the model generate intermediate reasoning tokens before the final answer to raise accuracy — from writing think step by step in the prompt to reasoning models trained to emit separate thinking tokens. Thinking tokens add latency and cost, so a thinking budget caps them; and since the chain of thought is not guaranteed correct, the result still needs verification.
When grown-ups say it
- Reasoning model
- A parrot trained to mutter first. It has the habit of laying thinking beans before the answer beans.
- Chain of thought
- The muttering, line by line. Write think step by step on the note and an ordinary parrot does it too.
- Thinking tokens
- Thinking beans. More of them than answer beans. Being beans, they cost money and time.
- Thinking budget
- How many beans it may mutter. Zero for easy questions, plenty for hard ones. The trainer sets it.
- Latency
- How long you wait for the answer. Grows with every thinking bean. → what an answer costs in time and money
- Cost
- The bill for thinking beans. Even beans you never see are billed. → the price of beans
- Self-verification
- Muttering, then checking again ✓. Raises the odds of being right. Not a guarantee — a wrong first step fools the check too.
- Hallucination
- Plausible muttering. Thoughts are continuation too, so they can be wrong. → the plausible parrot