Thanks to Will Shortz

For keeping us human—in crosswords at least for now

Cropped from Guardiansource

Will Shortz has just celebrated 30 years as the Crossword Editor of the New York Times. He remains the only person in the world whose college diploma lists the major as enigmatology, meaning the study of puzzles. He has also hosted a wordplay segment on NPR’s Weekend Edition Sunday since 1987.

Today we express our thanks to Shortz and his coterie for the proliferation of puzzles—while musing on implications for the intellectual future of humanity and non-humanity.

Shortz is only the fourth editor since the crossword was introduced in 1942. As Sunday’s NYT article saluting him says, this was done “as a way to offer relief to readers overwhelmed by war news.” We related last year that Samuel Fallows, in the introduction to the 1898 third edition of this thesaurus, plumped its use “for the solution of Cross Word puzzles.” We have featured the first editor, Margaret Farrar, and read into a guest puzzle by Bill Clinton in 2017.

The other two editors were Will Weng from 1969 to 1977 and Eugene Maleska from then until 1993. They introduced themes beyond the standard across/down pattern, such as Maleska’s Stepquotes in the Sunday NYT Magazine puzzles, but Shortz has ramped up that element. This Sunday’s puzzle, titled “Growth Spurts,” does so literally—as common phrases are extended vertically into jocular clue answers.

Computer-Aided Construction

I once constructed a 15×15 puzzle with theme answers typified by “BENCH BENCH” and “HARRY HARRY” but refrained from submitting it as a daily puzzle. I’ve never made one of the Sunday size. I appreciated the difficulty of making a puzzle by hand. In one puzzle for my college humor publication I made a one-box error (Major Bowes not Bowen) that was not caught by the editor. For the home computer that came with my set-up funds as new faculty in Buffalo, I bought (with my own money) a crossword construction program. What a neat concept, I thought—but I used it only a few times.

I imagine constructors—and editors—use them all the time. That brings some momentary musings:

• One can regard a crossword assistant as an advanced form of “autocomplete.”

• ChatGPT has often been dismissed as “glorified autocomplete”—note how Geoffrey Hinton mentioned and rebutted this two weeks ago.

• Despite their use for longer than Shortz’s thirty years, the puzzles—at least those I see—retain an unmistakably human element.

Alright, I first wrote “unmistakeably” but Google persuaded me otherwise. What may occasion this is another pattern I have noticed with the puzzles over the years: there are more clever clues. The crossword community has also encouraged novel words and phrases apart from the main theme answers. Novel by definition means not in a database of crossword usage that a construction assistant might store, nor in common dictionaries.

Still, these “human” attributes can be programmed into AI-powered crossword builders—even the clue conventions of British cryptic crosswords. Here is one recent attempt on auto-solving the latter.

TL; DP

What really catches me is the last quote from Shortz in the NYT article:

“Finally, I’ll say that crosswords are especially well-suited to the modern age, in which the world moves at lightning speed and our minds race from one thing to the next. A typical crossword has 76 or so answers, each on a different subject. The brain jumps from topic to topic to topic. Crosswords today feel more attuned to the times than ever.”

I’m teaching UB’s 4xx/5xx theory course again, and once again experiencing a cognitive style that differs from what I recall 30 years ago. I’ve had students confess to difficulty following proofs in lecture, such as NP-completeness of graph problems by component-design reductions. Numerous subsequent homework papers had too-short answers that gave a feeling of “TL; DP”: “Too Long—Didn’t Prove.” Rather than put a longer challenge problem on the next set, I posed more compartmentalized problems about mapping reductions.

In the meantime, Terence Tao has been discoursing about GPT-4 as a mathematical proof and writing assistant. Discussions in other forums expose gaps in reasoning and basic calculations—similar to examples in the presentation by Yejin Choi at TTIC which we featured last month. Looking to the (near) future, this thought springs up:

Insofar as the GPT architecture is based on finding the best next element, it may scale up well for generating mathematical proofs—especially what Juris Hartmanis and his students identified as “long and narrow” proofs.

This is a most-likely way that near-term AI may significantly outpace us. That social media is infamously unconducive to linear reasoned argumentation, while catching a greater portion of our brainwaves, can only accelerate this divide.

Open Problems

Before we wade further into deep AI matters, a simpler question is, can (Chat)GPT solve NYT crossword puzzles? Or British cryptics?