This site comes out of curiosity.
Most octopuses that have been tested have only one kind of light-sensing pigment in the eye, and their behaviour gives no sign that they tell colours apart. They match the colours around them anyway, closely enough to vanish against a rock. A language model's grasp of colour, and of things in general, is built from an enormous amount of secondhand material, not from having seen any of it. It will still describe a colour, and in my experience describe it more vividly than most people do.
Why? I do not have an answer yet. I am still looking.
Which brings me to mental imagery. One person says "the red apple in my head". Someone else says the same words back. The conversation goes fine. Between them they make an apple pie. Nobody checks, and nobody has ever needed to: it is an apple, what could there be to check?
Nothing goes wrong, so it does not occur to anyone that there is anything to check. But some people's apple arrives with the whole scene around it. For some it is nothing like a photograph: a blank, or words, or a wordless sense of it being there. Same sentence, different apples.
I am at one end of that range myself, and I only found out by accident. Until then I assumed everyone was like me.
Neither of the two below is an article. Each one moves you into a mode that is not yours, once. This is not the kind of thing you can find out with a single question: ask someone whether they can see the apple and they will say yes.
The imagery wall is one you do yourself, and you come out knowing which one you are, and that there are others. Almost is the model’s side: you type a sentence and watch, word by word, the candidates rise and only one take the seat.
The imagery wall
You imagine one apple. Then you change it three times, and you time yourself doing it. Then six versions of the same apple appear side by side, from a full picture down to nothing, and you pick the one that matches what was in your head.
Sources, and how each claim on it was checked
StartThe AI series
Every word in a sentence has only one seat.
Each time a model says a word, a whole row of candidates steps up at once; only one takes the seat. This page draws one sentence as terrain: a row per word, the candidates as hills, a circle on the seated one; the rest stand close by, unseen.
You can try it live: type a sentence of your own and watch it rise into hills, one word at a time.
The three built-in passages are real runs; the live one is generated once you press send, by asking the same model.
Record
None yet.
Information updated 2026.08