Shaan Singh [email protected]

I love, to a problematic degree, the New York Times Spelling Bee.

The goal is to find words using only the letters in the puzzle. Every word has to include the yellow letter somewhere in it, and 3 letter words are too short to count. Beyond that there are no rules. You can repeat letters, put them in whatever order you like, and so on. It quickly becomes addicting.

In the shower yesterday, I was thinking about the way I play the game. What is my brain doing when I play this game? I found that it was surprisingly hard to dissect the exact process going on in my brain. I also don’t know anything about brains. It was something like this:

• Look at the letters.
• A word will usually pop into my brain that kind of, sort of uses the letters. Example: “Cat”
• Riff off this word using sound associations (”bat”, “mat”, “sat”) or idea associations (”kitten”).
  • The riffing isn’t purely sound/idea associations. There is also this subtle background process running in my brain that is aware of the letters and trying to “push” the words to fit the puzzle.

An example of this process might be:

• Cat
• Mat
• Moat

I’ll annotate this example with my thought process:

• Cat: I glanced at the letters and immediately saw “Cat”, but it’s too short
• Mat: Rhymes with “Cat”, but doesn’t work because it’s too short
• Moat: I saw the “O” in the letters and realized I could go from “Mat” to “Moat”

From here I wondered: What if I could formalize different approaches to playing the game and simulate gameplay on my computer? Then I could measure which approach works the best, but even more interestingly, maybe I’ll learn something cool about cognition and/or GPT.

Method 1: Ask GPT to play like a human

The easiest way to simulate my approach to this game is to ask GPT to do it: