Using multi-LLM orchestration to decode complex puns, Sid Chatterjee’s passion project solved AI’s most stubbornly human word puzzle in under seven minutes.
When a computer program named Darmok completed the final puzzle of The Times Crossword Championship nearly three minutes faster than the human winner, it marked the fall of one of artificial intelligence’s longest-standing hurdles. For decades, British cryptic crosswords, dense with puns, visual tricks, double meanings, and allegorical wordplay, remained a uniquely human sanctuary—sparking debate over whether AI could ever truly master them. But for Prof. Siddhartha “Sid” Chatterjee, a computer scientist and faculty member at The University of Texas at Austin, breaking that barrier was never about commercial enterprise. It was the thrill of solving a stubborn puzzle.
“I am 100% devoted to teaching,” Chatterjee says of his role in UTCS, where he focuses on undergraduate courses in computer architecture. “This was a passion project because it was something that had been bugging me for a long time. Eventually I said, ‘Okay, let’s see what we can do with these LLMs.’” His work comes at a fascinating time for the craft, where human competitors—from young crossword solvers to seasoned grandmasters—are increasingly going head-to-head with artificial intelligence.
The Allegory of Darmok
To tackle a game built on deliberately deceptive wordplay, Chatterjee drew inspiration from science fiction. He named his system Darmok after an iconic episode of Star Trek: The Next Generation, in which the crew of the Enterprise must decipher an alien language constructed entirely of metaphors and allegories.
The project took root after an initial attempt last year ended in frustration when Darmok timed out after 18 minutes, leaving two clues unfilled. Recognizing that his early system was engaging in a loop of rapidly inserting and removing individual letters, Chatterjee went back to the drawing board to re-engineer its core operational logic.
Rather than relying on a single AI model, Chatterjee designed Darmok as an orchestrator that queries multiple frontier models in parallel—a setup he likens to legal AI tools. “It goes out to three or four different backends,” he explains. “It goes out to Gemini and Claude and GPT, with a prompt: ‘Here is a clue, and I want you to solve this clue and give me your reasoning for why you think your solution is correct. Give me your degree of confidence in that reasoning.’”
Parallel Logic and Grid Dynamics
When a puzzle begins, Darmok distributes all the clues simultaneously across its primary backends, Gemini and Claude, alongside GPT-5 and GPT-6 Astra. Once the backends respond, Darmok weighs the evidence locally before committing answers to the grid.
“What Darmok does is it feeds all of the clues off in parallel, and it waits to get the answers back to all of them,” Chatterjee says. “Then it looks at all the answers it gets, weighs the evidence, and says, ‘Okay, given that I’ve got all of this data, what am I willing to commit to the crossword at this point?’ This order-independent constraint orchestration is the algorithmic backbone of the redesigned system.
As intersecting letters are filled, Darmok generates candidate word lists and feeds them back into subsequent prompt rounds as hints. “Now all of a sudden you aren’t solving the remaining clues cold,” Chatterjee notes. “It says: ‘It might be this because of other information I have. Give some priority to this.’ And it continues in this way until it has fully solved the puzzle.”
For timed live competition, where accuracy and speed are equally important for winning, Chatterjee runs three identical copies side by side for redundancy in the face of live systems unreliability. “Whichever is the first copy to finish, that’s it,” he says. “You’ll see three identical puzzles being solved, but that’s the reason why.”
Piercing the Wolf Trap
Darmok’s strategic overhaul proved decisive during the championship final when it encountered 2 Down: “Wolf trap missing its lowest crosspiece? (5)”
The visual trick stumped three of the five human finalists, including 14-time defending champion Mark Goodliffe, who incorrectly entered SCARF. Darmok correctly identified SNARF (a slang term meaning to eat greedily, or “wolf”).
Chatterjee points to Gemini’s internal reasoning output, which evaluated the solution at 100 percent confidence: SNARE is a trap, and because the answer runs vertically down the grid, removing the lowest horizontal stroke of the capital letter E visually alters it into an F.
Becoming a Cryptic Clue
Darmok’s historic win did not go unnoticed by the world’s premier crossword authority. In a fitting meta-twist, The Times Puzzles Editor Mick Hodgkin honored Chatterjee’s creation by crafting a topical cryptic clue about Darmok for Times Radio listeners:
“AI not foxed by Times cryptic — respect! (10)”
The solution to the clue is ESTIMATION—where “respect” serves as the definition, built from an anagram (“foxed”) of “AI NOT” paired with an anagram (“cryptic”) of “TIMES”.
For an AI system born out of a desire to crack the hidden logic of human wordplay, becoming the subject of a Times cryptic clue brought the achievement full circle.
From Everest to Olympus Mons
Despite Darmok’s lightning-fast finish in under seven minutes, Chatterjee remains grounded about what the milestone represents. “A human solver brings one mind and a pencil,” he points out. “I bring four models, three copies—and rules that people spent months building. And a human’s wetware runs on 15 to 30 watts of power.”
For Chatterjee, the ultimate test of machine creativity is not solving puzzles, but setting them—a feat requiring smooth alignment between the surface grammar of standard English and the hidden grammar of “cryptic English”.
“If beating humans at cryptic crosswords was the Mount Everest challenge,” Chatterjee says, “I call this next one the Olympus Mons challenge.”
From the UT Austin Department of Computer Science website.
