Slam-Dunking the Madeleine

“Reading to the Endgame: A Novel Approach to Computer Chess” explores the possibility of two novels playing chess against one another by creating a program that assigns, for the black and white texts separately, to all sixty-four squares on the chessboard a combination of two letters—a “tuple”—based on the frequency that these tuples appear within the text. To make a move, the algorithm looks for the next tuple in the text that corresponds to a square containing a moveable piece, before finding the next tuple that corresponds to a square to which the selected piece can move. With this system, the algorithm must know only the rules of chess in order to conduct a game between any two texts.

One can attempt to manipulate this by working in reverse; that is, is it possible to write two texts that, when pitted against one another, conduct a controlled game of chess? I attempted this by writing two texts containing multiple different tuples to produce a position in which black checkmates on the second move (1. f3 e4 2. g4 Qh4#). Unfortunately, there was no way to form these tuples into sentences as there are too many variables that must be controlled (the frequency of every tuple, the order of the tuples, ensuring no new tuples are created when words are formed, and so on) for this to work.

Instead, I chose to start with a pre-existing text, but one that is already nonsensical: the German Dada artist Kurt Schwitters’s Ursonate, written between 1922 and 1932. When, as white, it plays a game against Proust’s Du côté de chez Swann, the game ends in a draw due to so-called threefold repetition. However, would it be possible to slightly alter the original text at the moment it is close to winning so as to ensure that it does?

With the help of a chess engine, I played through the game to look for positions where white has the possibility to checkmate black. I was able to find a forced (meaning that, regardless of what the opposition plays, it will be checkmate) mate in two moves that begins on just the twenty-first move, relatively early in the game. The mate itself is a very beautiful one; it involves a bold queen sacrifice on b6, forcing the knight to take it and creating a mating net around the black king, so that, when the white knight moves to b5 with check, the king has nowhere to run to and is mated. Instead of this brilliant maneuver, Schwitters begins by giving a check with the knight rather than the queen, missing the opportunity. This disappointing outcome can however be changed simply by adding four tuples, or eight letters into the Ursonate at the point where the algorithm was after the twentieth move, guiding the next two moves white plays to checkmate.

After adding the tuples that correspond to the necessary squares for checkmate, I attempted to run the doctored text against Proust’s novel, but it soon became clear the process was not so simple. The extra letters disrupted the tuple frequencies, meaning that the tuples for white now corresponded to completely different squares than before, thus changing the entire game. To restore the original tuple frequency, I pasted the Ursonate underneath my doctored version. After putting this updated text into the program, I was able obtain my checkmate on the twenty-second move.

Lastly, I “disguised” the tuples to appear more natural in the text. For instance, I changed the tuple “fu” to “füü” (as I discovered that the algorithm does not recognize the umlauts on letters) and “be” to “bee,” a term that occurs frequently in the original text. (Luckily, the extra letters did not change the tuple frequency again!) With this slight modification, I was able to discreetly alter Kurt Schwitters’s Ursonate to crush Proust’s Du côté de chez Swann.