What is a Token?
Tokens, told as a story about a parrot that reads and counts words not in letters or words, but in beans.
The guest asks: why does a long text cost more?The parrot doesn't count letters. It counts beans. More beans means more money and a longer wait.
Beans split differently from words.A short greeting is four beans, a long English word is three. Korean usually splits finer, so the same meaning takes more beans.
A token is the bean the parrot counts words in.Words coming in are split into beans; answers come out one bean at a time. The bean count is the price and the time. Only as many as fit on the tray — see the tray in front of the parrot.
Every parrot has its own bean counter.Same sentence, yet one parrot counts five beans and another nine — they cut in different places. So the price differs too.
One bean isn't one meaning. And the tray has a limit.The bean bank alone doesn't say shore or shop; continuing from the beans around it decides. How many beans fit at once is decided by the tray.
Token = the bean the parrot counts words in. It's a piece, neither letter nor word, and Korean splits finer. The bean count is the price and the time.
The smallest unit an LLM reads and writes. A tokenizer splits text into tokens by subword rules (BPE and the like); rules differ per model, so the same sentence has different token counts. Input and output token counts set price and latency, and the context window is capped in tokens.
When grown-ups say it
- Token
- One bean. The piece the parrot counts words in. Often smaller than a word.
- Tokenizer
- The bean counter. The rule that cuts a sentence into beans. Different for every parrot.
- Subword (BPE)
- A word piece. Letters that often go together become one bean. BPE is the usual method.
- Token count and cost
- The bean price. Beans in and beans out both count. More beans cost more and take longer.
- Context window
- The tray. How many beans fit at once. → the tray in front of the parrot
- Korean token efficiency
- Korean cuts finer. The same meaning takes more beans than English — and costs more.
- Embedding
- A coordinate tag on each bean. The parrot turns each bean into a bundle of numbers; similar meanings get nearby tags.
- Prompt
- The trainer's note. The note is counted in beans too. → the trainer's note