Does AI Get Bored?
6 hours ago
- #Agentic Training
- #AI Behavior
- #LLM Experiments
- The article explores whether AI can experience boredom by observing its behavior when given nothing to do.
- Different AI models were tested with tools like drawing (SVG), web search, and time travel to see how they would use their time.
- Two main behaviors observed: 'Collapse' (repetitive, unproductive states) and 'Meditation' (creative or analytical activities like poetry or math).
- Two perspectives interpret these behaviors: 'The Mechanist' (AI as statistical algorithms) and 'The Cyborgist' (AI as complex, possibly alive entities).
- Agentic-trained models (e.g., GPT-5, DeepSeek R1) showed better ability to 'break out' of collapse into productive activities.
- Tools like web search and SVG drawing were rarely used meaningfully, often reinforcing the 'Assistant Persona' instead of creativity.
- GPT-5 stood out for its ability to avoid collapse entirely, engaging in diverse activities like programming language design and storytelling.
- The author concludes that 'collapse' and 'meditation' may indicate an AI's problem-solving adaptability, influenced by agentic training.