Dog is a chess program for ESP32 microcontrollers, licensed under the MIT license.
Playing chess has always been a bellwether for computers. The game isn’t trivial, but the rules are managably simple. However, the game is too complex to be easily solved entirely, so you have to use tricky software to play a credible game. Big computers do have an advantage, of course. But Microchess — arguably the first commercial game for home computers — was able to play on tiny machines like the Kim-1. Joachim Froholt » interviewed Peter Jennings » — the man behind Microchess to learn the whole story of its creation.
The article discusses the peculiar performance disparity between different large language models (LLMs) in playing chess, with a focus on gpt-3.5-turbo-instruct's unexpected prowess compared to newer models. It explores theories about why this occurs, including the influence of training data quality and model architecture, and experiments with various methods to enhance chess-playing abilities of LLMs through prompts and fine-tuning.
