The describer calls your BMX rider "a biker in a skate park". You fix it. Three hundred pictures later it says "biker" again, which is the kind of thing that makes people distrust machines and, occasionally, people.

It can't be retrained; it's Apple's on-device model, and it arrives exactly as it is. But a model doesn't need to learn to be told. Every request carries a short brief with room for notes: 1,200 characters, roughly a postcard. The notes are, in order, the one line you wrote for the project or album, the words you prefer for pictures like this one, the picture's keywords, the words you used on three lookalikes, and whatever the pictures taken within ten minutes already say.

The preferred words come from watching you edit. Change "a biker rides a ramp" to "a BMX rider rides a ramp" and the difference is a lesson: biker, BMX rider. Rewriting the sentence teaches nothing; nor does swapping "a" for "the", adding a word or removing one. Only a swap counts, and swapping it back forgets it.

A lesson is kept on a short lead. Told "Toyota, not Citroen" for every picture, an early build tagged a desk "toyota". So a lesson holds only where it was learned: the same picture, a lookalike, the same place, a shared keyword. And the example captions are given as words, not sentences, because one sentence about a bowl put a bowl into a third of the replies, and no picture had one.

The model never learns a thing. It just keeps getting better notes, which is how most of us get through the day.
