The biggest chess game ever

00:00:00.000 | This is a game of chess played by two AI engines each far better at chess than any human on earth.

00:00:06.720 | The question I had is what happens when you let these machines battle it out

00:00:12.080 | on an infinite chessboard with infinite pieces.

00:00:15.120 | Each move calculated and performed on an 8x8 subset that in isolation is a totally legal

00:00:24.480 | position with two kings but in the big picture is just a tiny subset of a much bigger game.

00:00:30.960 | Where each checkmate destroys the king and the deserted pieces move on in search of another

00:00:38.160 | neighboring victim. Pieces get captured and sacrificed in other words the usual game of chess

00:00:44.800 | but on a much bigger board. The play is extended arbitrarily outwards but here we focus on just

00:00:53.440 | this subset of about 6,000 squares. On the bottom is the counter showing the number of kings and

00:00:59.040 | checkmate victories. Let me mention some details. The boards are initialized with the middle game

00:01:05.440 | position for one of 30,000 famous grandmaster games including Magnus Carlsen, Bobby Fischer,

00:01:11.600 | Gasparo, Spassky, Tal, Karpov and so on. The selection of which 8x8 subset to use for computing

00:01:18.560 | the next move is made randomly among all the boards that have legal chess positions. Optimizing

00:01:24.400 | the selection process is something that could be formulated as a reinforcement learning problem

00:01:30.000 | and would make this meta game of infinite chess very interesting. In fact there are fascinating

00:01:36.240 | questions here about the generalizability of AI engines far beyond the scale and the constraints

00:01:41.600 | of the game they were trained on. Chess is just a game but when you start to remove the constraints

00:01:50.400 | and increase the scale it becomes something else. Something more like the game of life

00:01:55.520 | played on an infinite chess board.

00:02:39.120 | [Subscribe]