A test that image generators failed for two years: asked for a glass of wine filled to the very top, they returned the usual third, because almost every photograph of wine shows a normal pour. Popularised by a philosophy YouTuber's video in early 2025 as an example of models reproducing what they have seen rather than what was asked. Newer models pass it, mostly. The record's entry for the pour that would not go up.
Seen in March 2025 when a new image model managed the brim on the first try and a thousand people posted it like a moon landing. A week later they were asking for it to overflow.
I tried it on four models in an afternoon in 2025 and got four elegant, sensible pours. The machines have better manners than I do at a wedding.
The record notes the wine glass belongs with the salmon and the clock: cases where the training data has one overwhelming version of a thing and the model cannot imagine the other. The glass is the clearest because everyone has overfilled one.
What does it mean? Write it the way you would say it out loud.
