In other words, they did not even need a prompt. You could probably skip feeding the paragraph and just told "generate visuals for the opening paragraph of the

In other words, they did not even need a prompt. You could probably skip feeding the paragraph and just told "generate visuals for the opening paragraph of the LotR", and the LLM would successfully do it as it has a very good idea of what they are from its training data. This is a bigger difference to the pelican than simple reproducibility steps. The pelican is intentionally esoteric, and thus open ended. The LotR is mundane and has a "correct" answer, aka, copy the movie. It makes it a really awful test of capabilities. The pelican isn't a slop test. This crap is.