I.E. A.I.
« previous post | next post »
In an update to "Morpho-phonologically AI", I wrote
Ironically, since this puzzle was vocalically inspired by the term "AI" , I'm guessing that current AI systems are not very good at solving (or creating) puzzles like this. I'll give it a try later today.
But it seems that I was wrong.
From GPT-5;
"Iced tea" is cold, not frozen — but still, pretty good. It would be a bit harder if a restriction to monosyllables were added, but GPT-5 handles that as well:
My pessimistic prediction was prompted in part by things I've read about (previous) ChatGPT systems' problems with crossword puzzle clues — but maybe they're improving?
Kenny Easwaran said,
August 13, 2025 @ 1:21 am
The link you posted at the bottom is from April 2024 – at that point, people were aware that if you gave the model the phrase "let's think step by step", it would list the steps in some hypothetical reasoning, which would then give it better background for the later steps, and would lead it to useful answers more often than if it just had to predict the answer with no intermediate words. In September of last year, when ChatGPT introduced o1, that was the first public model that did that sort of thing by default, and also had done some practice at figuring out what sorts of steps more often get to useful answers. By February or March of this year, all of the other major labs had "reasoning models" as part of their default model and ChatGPT 5 is the first time OpenAI is bringing the "reasoning models" into their default model.
Coby said,
August 13, 2025 @ 3:15 pm
In my experience, AI (at least Google's version) is still totally hopeless in proposing solutions to cryptic crosswords.
Circeus said,
August 14, 2025 @ 10:38 am
I've tried them occasionally for "categories" type puzzles where you have like, 16 words and have to figure out four groups of 4 words. Because the answer are just as likely to be thematic ("one word movie titles"), phrases ("words that come before 'card'") or metalinguistic ("Homonyms of famous brands", "words with several pronunciations"), AI tends to whiff really badly.