I accidentally omitted a sentence. I tested it on DallE mini as well. It also behaved consistently there, but that version gave me mostly Beetles with a few bike races mixed in. I also played slightly with the nonsense words to see if perhaps something I picked was too close to a real word. The revised prompt gave me doctors on DallE mini and letter tiles / notebook pages in DallE, with the mini-version producing two noteworthy images that exhibit the "name one thing in this photo" effect.

https://i.dailymail.co.uk/1s/2019/04/25/14/12707994-6959547-...

I'm not particularly surprised that official DALL-E & unofficial DallE-Mini went different directions with the same input, as they're different-sized models with different training sets.

Below a certain level of complexity, perhaps only "vaguely like beetle or bicycle/race names" arbitrarily survived all the other competing influences on the model's internal weights. Above, much more subtle patterns – like many more latinate word roots – might start to survive.

Compare also what I linked in another thread branch – https://twitter.com/gojomo/status/1540095089615089665 – where simply growing the size of a similar (non-public Google PARTI) model leads to a phase-change in its ability to render meaningful human text.