OpenAI’s chat GPT O3 model defeated AI Chess Tournament organized by Kagal to defeat Grocke 4 of Allen Musk’s company XAI. In this three-day competition, large language models (LLMs) of the general purpose of many companies were face to face. The special thing was that none of these models were specially designed to play chess.
Tournament format and participant
Eight models participated in the competition, including OpenAI, XAI, Google, Anthropic and Chinese companies Deepcic and Moonshot AI models. Standard chess rules were implemented in the game, but the challenge was that these general AI models had to show their ability in strategy and moves. According to the BBC report, Google’s Gemini model finished third and defeated another model of OpenAI.
What happened in the final?
Groke 4 took an edge at the beginning of the tournament, but the finals saw many strategic mistakes in her game – including losing the queen repeatedly. ChatjPT O3 took advantage of these mistakes and won a decisive victory. “Until the semi -finals, it seemed that no one could stop Groke 4,” said Pedro Pinhata, author of Chess.com, but in the final, his game shattered under pressure. Grandmaster Hikaru Nakamura, who was making live commentary, also admitted that Groke made constant mistakes, while ChatjPT performed a stable game.
Alone Musk’s response
After the defeat, Elon Musk did not take this result much seriously and said that the strong performance of Groke was just a “side effect” because Xai had almost not done any work on chess. This result makes the competition running between the XAI of OpenAI and Musk even more public. Chess has long been a medium to test AI’s abilities. Even earlier, special AI systems like Google Deepmind created history by defeating human champions, but this tournament was special because it was tested for general LLM, not a particular chess engine.
What does it mean?
This results indicate that large language models can perform differently in sports with structured and competitive environment. The stable performance of the chat GGPT O3 shows that some models may continuously perform strategic games, while the Grocke 4’s necklace shows that the results may be unstable under pressure.
