Most chess engines search through a tree of branching possible futures to pick their next move. But in the NeurIPS 2024 paper Amortized Planning with Large-Scale Transformers: A Case Study on Chess, Google DeepMind researchers partly distill the gameplay of the chess engine Stockfish 16 into a series of transformer models that do not search the game tree. This page runs their 9M-parameter, ~2000 Elo model in the browser so that you can play against it.
CHOOSE YOUR SIDE
Instead of searching through branching possible futures where the AI moves, then you move, then the AI moves again, and so on, this model thinks just one move ahead: it scores all of its legal moves and chooses the one it assigned the highest score. You could already do that by removing the search operation from existing chess engines, but what makes this project different is that the researchers tried to distill the way Stockfish 16 moved based on its search into transformer models without any game tree search component. From one point of view, the experiment failed: these models do not play nearly as well as Stockfish 16. But they do play pretty well by human standards, and they are interesting artifacts one of which is small enough to run in the browser.