Rich Felker@tobtobxx @GossiTheDog My naive guess is that this is because it's not programmed to vomit the best low-correlation candidate when none of them correlate well with the corpus, but instead a statement of uncertainty and needing more context.
The industy's LLMs won't do this, because being confidently wrong (and confident in general) is how they bamboozle users.
All GNU social JP content and data are available under the