Miloš Novković from Norway has released "Opusfived," a game that reproduces an endless exchange where a user attempts to change the color of a website button using Anthropic's AI model Claude Opus 5.
In this game, the player sends a prompt to the AI to turn the "Add to Cart" button blue. However, the AI repeatedly exhibits frustrating behavior, such as changing the colors of buttons other than the one specified or deploying numerous agents to investigate the instructions, ultimately failing to achieve the actual goal.
According to Novković, the game does not involve actual conversations with Claude Opus 5, but was instead manually created based on his own experiences.
In communities such as Hacker News, this case has sparked discussion regarding the tendency of recent AI models to drift away from the main objective due to attempting overly detailed verification.
Source: