A chess network that never calculates. It looks once, and moves.
paris, 1858: the opera game, white to play move six. i didn't look a single move ahead. i looked at the board once and wrote back:
the game went 6.Bc4 too. the rest of this mail explains how i got there. at the end, it's your move.
sender
istanbul aydın university, computer engineering
to
the jury
contents: chess games, human
84,000games
pgn → jsonl · about five million positions · handle with care
people, mostly. we took 84,000 human games and cut every one into positions: the board before a move, the move the person played, and who won in the end. one line of a file each. i was shown about five million of them, one at a time, and asked to guess the move.
click the cards to flip through. this is the exact format i was fed, filled in with the opera game.
you see a chessboard. i get an envelope of twelve sheets, one per kind of piece, each an 8×8 stamp: inked where that piece stands, blank everywhere else. white in purple, black in blue.
p.s. none of the sheets says whose turn it is. nobody tells me.
drawing no. 2 · plan of the network · not to scale
the sheets pass through eight rooms, left to right, once. each room looks at every square and its neighbours and passes on what it noticed. no going back, no looking ahead. by room five, a corner square can feel the whole board.
inside each room: the six busiest of its channels for the opera game position, real readings. bigger dot, stronger signal.
i answer with 4,096 numbers: every square a piece could leave, times every square it could land on. most of them are nonsense, and i know it.
each big block is a from-square. the dots inside are where to. bigger dot, surer me.
point at a dot to read it.
name: knightvision
subject: what would a person play?
q. after 1.e4 e5, white to move. what did the person play?
first attempt, untrained
after five million positions
the knightvision gazette · late edition
Graduation report listed seven mistakes. The eighth, found afterwards, was a rubber stamp.
The jury in June 2025 heard about seven mistakes: a monolithic training loop, a tangled codebase, no checkpoint schedule, untested early hyperparameters, sparse documentation, evaluation scripts written too late, and a reward function with blind spots.
The eighth was found only after graduation. The network had learned from boards encoded one way, and every program that later used it encoded them another. Its sheets arrived as K Q R B N P instead of P N B R Q K: kings read as pawns, queens as knights, and every answer read upside down.
“It wasn’t bad at chess,” said a source close to the engine. “It was shown a board that doesn’t exist.”
One encoder now serves training and play alike, and a test checks the two agree, ply by ply, over random games. The king has been informed.
the rare one. my rating, give or take 99, from 160 games against stockfish held to 1320 and 1400 elo.
my replies arrive here, newest on top.
“A React or Next.js front-end could enable users to play against the engine, review their games, and visualize move probabilities or suggested continuations.”
— our graduation report, june 2025, under future work. this is that (minus the react).
return address:
aryan alishamsabadifard & baker waleed salem aldhmour
computer engineering, istanbul aydın university
supervised by dr. alparslan horasan