0XP
All concepts
Concept 01 of 090/3 mastered

Tokens

Why the model never sees letters

Before a model reads a single word of your prompt, your text is chopped into tokens, fragments that are usually smaller than a word. Almost everything else (cost, limits, spelling mistakes) traces back to this step.

It never sees letters
Page 01 / 05

It never sees letters

Before a model reads a single word of what you wrote, your text is cut into pieces called tokens, and each piece is swapped for a number. The letters are gone from that point on.

A token is usually a bit smaller than a word. Common words get one each — including the space in front of them. Rarer words get shattered into fragments: 'unbelievably' might arrive as un + believ + ably. Invented words fall apart completely.

The strip along the bottom is the characters. Everything above it is what the tokenizer did to them.

Where the breaks fall

The strip along the bottom is the characters you typed. Everything above it is what the tokenizer did to them — the taller a block, the more characters it swallowed whole.

CanyousummarisethequarterlyreportbeforeFriday'smeeting63 CHARACTERS13 TOKENS · 4.8 CHARACTERS EACH
What this shows
5 of these 13 pieces were not on the list and had to be assembled out of smaller ones. That is where an uneven token count comes from.
Think of it like this

Like reading a language written only in syllable blocks

Imagine every text you receive is printed on tiles, one syllable-ish fragment per tile. You can read fluently, write beautifully, and discuss anything. Now someone asks you how many times the letter R appears in a word. You would have to prise the tile apart, and you have never once been allowed to do that. You would guess, and you would sound confident doing it.

CheckpointPick one

Why do these models famously miscount the letters in a word?