Crossword Help Forum
Forum Rules

John Green passed the torch to Neil Aspland at the end of last year

Re AI, I occasionally amuse myself by asking AIs to solve cryptic clues. It’s a sort of contemporary bear-baiting; as the task is much better suited to human fuzzy logic than current machine logic. But when they get better at it - and they surely will - I will start to be afraid!

They are not far away, jack, as shown by this report from this year’s Times Crossword Championship by puzzles editor Mick Hodgkin, in case you haven’t seen it:

“There is no stopping Mark Goodliffe, it seems, but the dawn of AI means The Times crossword champion was pushed to his limits by specially trained cryptic crossword software.

Goodliffe, 60, a former financial director, retained his title on Saturday (his 14th victory), solving the final puzzle in under nine minutes, more than four minutes ahead of his nearest rival.

But also pitting its wits (if that is the right word) against the puzzles was a specialist artificial intelligence program that has come closer than any other to cracking the cryptic crossword mindset.

Professor Siddhartha Chatterjee, a computer scientist at the University of Texas, brought his creation, Darmok, over on his laptop for the event after a summer of intensive training.

Darmok solved two of the three semi-final puzzles in an impressive seven minutes but came unstuck on the third. Faced with the final puzzle, the program put up a performance that would have earned a human contestant fourth place but gave up after 18 minutes with one clue left. This was it:

Large drops in circuits elevated battery discharge (5)

Chatterjee, who created the program only four months ago with Nick Tomlin of New York University, said it was based on the latest ChatGPT model, GPT-5 ,but with further prompts. He described it as “a reasoning model” which goes through a chain of logic to get to the answer.

Speaking after the event, Chatterjee said that they would keep refining the model and hope to return next year.

“Teaching an AI program to solve cryptic crosswords was one of those things I always thought must be impossible, so I had to try it,” he said.

Enthusiasts hope cryptic crosswords will prove to be one of the last bastions of human ingenuity that cannot be mastered by AI.

Many clues work like equations, with certain words indicating an anagram or that one word goes inside another. But solvers also need to exercise a degree of lateral thinking, and unravel punning cryptic definitions that are harder to reduce to crackable code.

However, every year the crossword’s equivalent of Deep Blue is coming closer. IBM’s supercomputer defeated the chess grandmaster Garry Kasparov in 1997, two years before Mark Goodliffe won his first crossword championship.

Now the question is who can beat Goodliffe — will it be a fellow human or an AI challenger?”

Thanks, Quisling. I hadn’t seen the report and I was very interested to read it…slightly chilling!

I made more of a meal of this than I should have, as I copied two figures down wrong. When I went back to it, it was plain sailing, and all the more satisfying for finding my blunder.

Just to add to the comments about AI, for me it's an irrelevance. Surely one of the pleasures of these puzzles is completing them with the minimum of resources. I just used a calculator, pencil and a piece of paper (hence my mistake). In doing so I learnt lots about prime numbers and powers, and also to write more clearly!

By the way, one quibble about the instruction: I initially interpreted 'the cells I landed on' as including the head and foot of each snake and ladder, just as in Monopoly you land on a Chance and then have to go to jail or somewhere. Adding 'after each move' would have been clearer.

Well done, Mack.

I think that the quibble is unfair. The preamble's example removes any ambiguity.

Good point. I must remember to look at the examples more carefully. Another thing learnt.

An old fossil is fascinated by all the chatAI stuff. The prime-raised-to-a-power clues took less than 10 minutes with a calculator to derive the very small number of possibles (powers are very restricted: only two are possible). Have we, as human beings, lost the ability to do simple maths? (That's not meant to be critical of solvers who use techy things, though it may sound it. It's really a plea lest children are no longer taught how to do straightforward arithmetical things, and thus to fail to understand the bones of mathematics.)

Rant over. Now for Ms Reeves and her Pandoric box of tricks.

I have just test-solved a puzzle that is in Rhombus style. Three 3-digit numbers between them contain all of the non-zero digits such that various conditions are satisfied.

For the first time ever I used the AI feature in Google Chrome which produced all the sets of results for each clue. They were all correct. It would appear that it uses Python and a brute force approach.

The puzzle, for me, would have been impossible to solve without it.

I don't think this is the end for this type of puzzle in The Listener, at least, since the puzzle must be doable with a standard scientific calculator and have a pathway that doesn't involve computer programming.

This not-quite-a-fossil can't find anything interesting about AI. It has yet to surprise me with anything but speed. But I agree with your observation about mathematics. I did this puzzle three times just to improve my understanding of the potential simplicity of the solving and what I might be missing by having to rely on a spreadsheet. I was astonished. Worth looking back to the very first mathematical Listener crossword and imagining how on earth that received the most successful completions to date. I remember a statistics professor who said that having to write out regression by hand made it clear that adding any additional variable increased r.