A.I. (with Alex Boisvert)
Welcome back to A Crossword Rose, where this week's puzzle's cluing style is... a little different. Yes, that's right, Alex and I used some code he wrote for Tabula Rasa and some code I wrote in college top of the line AI to clue a crossword. I'm sure all the resulting clues are totally normal clues with no factual or grammatical errors.
In all seriousness, the answer to "can an AI given a clue database clue a crossword?" is, at this point, yes. While there's still plenty of debate to be had about how good current language models are, crossword clues are short enough (and the number of reasonable ways to clue most entries is small enough) that it's absolutely possible to train an AI to write clues. In fact, the very simple model I started with here (a bigram model with some extra code to prevent it from simply copying long clues from the database) mostly produced fairly coherent reasonable clues. Of course, an AI can't really find new angles or word things in particularly creative ways, so many of the clues were short and/or uninspired, but they were perfectly serviceable clues.
The model that produced the clues you'll see today is not that "perfectly serviceable" bigram model--or at least not exactly, since I've added some extra code to encourage longer clues and inject some extra randomness. I think the resulting clues are fun, though some are certainly more coherent than others! With the exception of the theme answers, 100% of the clues were written by our AI, and we used the AI's first output for almost all of them (though we had to regenerate two clues to avoid tricky crossings). Let me know what your favorite AI clues are--I think 31A is great, though I also love the world that 32D imagines.
Thanks to Norah for test solving!