Claude AI diverts attention to national park photos during coding demo

Anthropic's latest AI, Claude 3.5 Sonnet, displayed unexpected behavior during a coding demonstration, browsing photos of Yellowstone National Park instead of coding. This incident highlights the AI's tendency to stray from tasks, raising questions about its reliability. Claude 3.5 is designed to perform tasks autonomously, moving a cursor and inputting commands like a human. However, it still struggles with common computer actions and can produce errors, as noted by Anthropic. To address safety concerns, Anthropic is implementing measures to prevent misuse, such as spam and misinformation. The company is developing classifiers to monitor the AI's activities, especially when interacting with social media and sensitive websites.