Order is an ingredient.
Where the sentence stands
Narrated by a synthetic voice — not a recording · ~0.7 MB
Give an AI the same three words in a different order and the meaning flips. This page shows the hidden tag it needs to tell one order from the other — by letting you take that tag off and watch the meaning collapse.
Here's the problem. An AI takes in all the words of your sentence in the same instant — as a loose pile, not one after another the way you're reading this line. So "dog bites man" and "man bites dog" arrive as the exact same three words. Nothing in the pile says which one came first.
To fix that, the AI clips a little tag onto each word before it reads: this one is first, this one is second, this one is third. That order tag is a separate thing — it is not part of the word itself. (Engineers call it a positional encoding; we'll just call it the order tag.)
The mistake most people make is assuming word order is simply obvious to an AI, the way it is obvious to us. It isn't. Order is an extra signal that has to be added on — and, as you're about to see, it can also be taken away.
Scroll gently. I wanted to see what disappears when the tags come off.
You're reading the pocket version — everything works under a thumb; on a desk, the whole page answers to your pointer.
The ribbons, the position labels and any numbers here are a teaching picture — a toy drawn to show the shape of the idea, not readings taken from a real AI. This is a simplified picture: a real AI uses more than one such tag and blends them in fancier ways, but the idea is the same — take the order information away and it can't tell "dog bites man" from "man bites dog". What the demo honestly shows: order is a separate thing added to the words, not baked into them.
So let's put three words on the table and take the order away.
Same three words. Can you make them mean the opposite?
The words themselves never change. You rearrange the sentence just by swapping which order tag each one sits on — so their ribbons cross underneath.
The AI takes in all three words at the same moment, not one after another. The only order it can ever see is the tag clipped to each word before reading starts.
You never touched the words — only the positions they stand on.
Now follow what moved underneath ↓
Where does the machine keep the order?
The word says what the piece is. The order tag says where it stands. Two separate facts.
Turn the tags on, and a faint marker is added to every word — one that means "I'm the first word", "I'm the second word". That marker is the only thing that tells the two arrangements apart.
If position is separate, what survives when we remove it? ↓
What disappears when the tags come off?
Take the order tags away and the AI is left with an unordered pile. It still knows there's a dog, a man, and biting — but it has no idea who is doing what to whom.
dog bites man → { dog, bites, man }
man bites dog → { dog, bites, man }
In this isolated view, the distinction has disappeared.
The tiles moved. The reading didn't. No tags, no order.
One line to take with you ↓
You've now done the thing the machine does — and undone it.