When the IBM supercomputer Deep Blue beat Garry Kasparov at chess in 1997, it was an early victory for artificial intelligence over the human mind, and the culmination of decades of attempts. Almost 30 years later, a computer program has finally outdone humans at solving cryptic crosswords. Earlier this month, a model called Darmok solved the final puzzle of the Times’s annual cryptic crossword competition about three minutes faster than Liam Hughes, 29, who became the contest’s youngest (human) winner.
After Deep Blue’s victory, the New York Times’s puzzle editor, Will Shortz, set engineers a challenge. I don’t think so,” he wrote. But cryptics are much trickier, and as recently as 2024 my colleague Alan Connor was asking: “Will the cryptic crossword remain the last thing AI can’t conquer?” Dr Sid Chatterjee, a computer scientist at the University of Texas in Austin, thought not, and in 2025 set out to prove it.
He began work on Darmok, named after a Star Trek episode in which the crew attempt to converse with a species that speaks only in allegories – an experience akin to solving a cryptic. Darmok’s first tilt at the Times championship ended in failure when it gave up on the final puzzle after 18 minutes. Its triumph this year was partly as a result of successfully solving a clue that tripped up three of the five finalists, including 14-time champion Mark Goodliffe: Wolf trap missing its lowest crosspiece? (5)” The letters filled in as part of other clues gave the pattern S_A_F.
Both SCARF and SNARF mean “to eat greedily” (synonyms for “wolf”) but the rest of the wordplay relied on a visual element. While one might expect visual clues like these to be more difficult for an AI, Chatterjee said he couldn’t detect a pattern in the types of clues Darmok struggles most with. It does find puzzles by John Henderson (known to Guardian readers as Enigmatist) harder than most, though.
If you’d like to challenge yourself against Darmok, it solved this Fed crossword from last week in four minutes and 10 seconds. Darmok is based on four AI models: Gemini 3.1 Pro, Claude Fable 5, GPT-5 and GPT-6 Astra. The solving is primarily done by the first two, with the others joining in if needed.
Chatterjee attributed Darmok’s improvement this year to the extra power behind these frontier models, as well as some of his refinements, such as making Darmok wait until it had attempted all the clues before inserting answers. The next frontier for AI and cryptic crosswords – setting a satisfying puzzle – will be an even bigger jump. It’s roughly 45% and 42% (5) From Bluebird in the Independent.
Extract — continue reading at the source.
