Post

I wonder how it would fare against a human player of any skill. There are "good" moves, "bad" moves, and moves that make no sense if you understand the common strategies, so holy crap this guy must be a genius. Could it fool a wary player into thinking it's better than it is?

Try sorting by arranging piece locations on two separate matrices for each the letters and numbers then multiplying them and throwing a dart at the resulting combined matrix It may not result in legal moves but you will never lose the element of surprise.

Obviously white is in devastatingly bad position here. Next turn is unknown, but the obvious black is Qxf3. If I was white, Bg2 sacrifice then Pe5 or else its over. You could also make the case for Qd2 to make the attack really expensive and hope you can unpack that corner.

They should clearly choose a better ChoosePivot subroutine to enhance the performance of the engine. My suggestion is to use some kind of randomized element to so the "expected" worst-case performance is better.

This begs the question of how the board state ever got here . . . By this sorting method the knight is going to move almost all the time, and typically that's going to be moving the knight between two spots indefinitely

It's about legal moves, so the first few will be pawns as there are 4 legal knight moves at the start of the match (and the top of the list) but 16 legal pawn moves

Title text: "It struggles a little with complex positions, like when there are an even number of moves and it has to round down, but when run against itself it's capable of finding some novelties. At one point I saw six knights on the board at once; Stockfish rarely exceeds four."

The alt text strongly reminded me of the time when someone made ChatGPT play against stockfish. That game was wild. Involved several resurrections of pieces and heavy teleportation