@CarlMuckenhoupt@mastodon.social
Post #1886368
2026-04-25 02:03 UTC
I've just been experimentally feeding cryptic crossword clues to ChatGPT, including both published ones and ones that I made up and whose solutions are therefore guaranteed to not be in its training data. I find the results interesting. ChatGPT is TERRIBLE at finding the answers, but actually really good at figuring out the wordplay once you tell it what the answer is.
Replies (1)
-
@CarlMuckenhoupt@mastodon.social 2026-04-25 02:10
My understanding is that LLM training data is tokenized in a way that means that the model doesn't actually have any record of how words are spelled. So it's no surprise that it comes up with answers that don't fit the enumeration, and are based on wordplay that simply doesn't work, like anagramming a word into something with completely different letters. But I'm quite surprised that this disadvantage completely vanishes when it knows the desired outcome.