Game #1
Translate a sentence for a machine — then translate its reply back.
A language model never sees words. It only ever sees numbers. You're about to feel why that matters.
Step 1 · Words → numbers
The machine can't read the sentence — swap every piece for its number.
Step 2 · Inside the machine
IN
processing…
OUT
Symbolic view — but the point is exact: inside, only numbers move. Never words.
Step 3 · Numbers → words
The machine replied — but only in numbers. Turn them back into words.
The reply
To the machine, "strawberry" was just 4842 8123. The r's were never there to count.
You can't count what you can't see.
The ID numbers are illustrative — they show how tokenisation works, not any specific model's real vocabulary.