AI Engines on an Infinite Chessboard: Exploring Generalizability and Emergent Complexity

The podcast introduces a fascinating experiment involving two advanced AI chess engines, each superior to any human player, engaged in a game on an "infinite chessboard." This setup drastically alters the traditional rules of chess by removing scale and constraint limitations. Instead of a fixed 8x8 board, the game extends arbitrarily outwards, with moves calculated on dynamically selected 8x8 subsets that are locally legal but contribute to a much larger, evolving meta-game. The core inquiry revolves around observing the behavior and capabilities of AI when faced with such unprecedented complexity and scale.

A key distinction of this infinite chess variant is its dynamic nature: each checkmate results in the destruction of a king, and the remaining pieces continue their search for new targets, creating a continuous, evolving conflict across the vast board. The game boards are initialized using middle-game positions from a database of 30,000 famous grandmaster games, including those of Magnus Carlsen, Bobby Fischer, and Garry Kasparov. However, the selection of which 8x8 subset to use for computing the next move is made randomly among all legal positions, introducing an element of stochasticity and further complexity beyond the deterministic nature of typical chess AI.

The host suggests that optimizing this subset selection process could be formulated as a reinforcement learning problem, highlighting a potential avenue for future AI research and development. This experiment is not merely about determining a winner in a conventional sense, but rather about pushing the boundaries of AI performance and understanding how these intelligent systems adapt and strategize in environments far removed from their original training parameters. It offers practical insights into the design of more robust and adaptable artificial intelligences.

The broader implications of this experiment touch upon the critical concept of AI generalizability, questioning how well AI engines can perform and adapt beyond the specific constraints and scale of the games they were trained on. The host draws a compelling parallel to "the game of life," suggesting that by removing constraints and increasing scale, a simple game like chess transforms into something akin to a complex system with emergent properties, offering profound insights into artificial intelligence, complexity theory, and the fundamental nature of computation and adaptive systems.

this is a game of chess played by two ai engines each far better at chess than any human on earth

what happens when you let these machines battle it out on an infinite chessboard with infinite pieces

each move calculated and performed on an 8x8 subset that in isolation is a totally legal position with two kings but in the big picture is just a tiny subset of a much bigger game

where each checkmate destroys the king and the deserted pieces move on in search of another neighboring victim

the play is extended arbitrarily outwards but here we focus on just this subset of about 6 000 squares

The boards are initialized with the middle game position from one of 30 000 famous grandmaster games

optimizing the selection process something that could be formulated as a reinforcement learning problem

there are fascinating questions here about the generalizability of ai engines far beyond the scale and the constraints of the game they were trained on

chess is just a game but when you start to remove the constraints and increase the scale it becomes something else something more like the game of life played on an infinite chessboard