When a simple 4-kilobyte game from the 1970s checkmates today’s advanced artificial intelligence, it might be time to rethink the future of technology.
Robert Caruso, an engineer at Citrix, recently conducted an unexpected experiment and arrived at some surprising results. Using the Stella emulator, he ran Atari Video Chess, a game released in 1979 for the Atari 2600, and pitted two modern AI models—ChatGPT and Microsoft Copilot—against it. The outcome? A 46-year-old game with just 4 KB of data managed to defeat both contenders.
The challenge began with a conversation between Caruso and ChatGPT about the differences between chess engines like Stockfish and AlphaZero. During the exchange, ChatGPT confidently claimed it could easily beat Atari’s Video Chess. However, in practice, the chatbot couldn’t even track the positions of the pieces correctly during gameplay. Despite Caruso’s continuous guidance, the model was defeated at the beginner level after a 90-minute match.
Caruso repeated the same challenge with Microsoft Copilot. This model was more confident, claiming it could follow the chessboard accurately. Yet within the first seven moves, it lost two pawns, a bishop, and a knight, and even suggested moving its queen directly into the path of the Atari queen. Ultimately, this game also ended in failure.
Caruso’s simple experiment highlights that despite all the marketing hype around AI capabilities, these models still struggle with fundamental concepts—such as understanding a basic chessboard layout. Contrary to corporate claims about replacing humans in complex professions, these models still lack true comprehension, abstract reasoning, and stable memory. If a large language model can’t distinguish between a rook and a bishop, how can we trust it in critical areas like medical data or energy systems?
