Does AI Get Bored? (timkellogg.me)

A researcher gave various LLMs a simulated “10 hours” of idle time (mapped to a token budget — e.g., 100,000 tokens = 10 hours) and minimal human prompting, sometimes equipping models with simple tools (draw_svg, search+fetch, a cautionary time_travel). Across models and runs the author observed three recurring behaviors: collapse (long stretches of near-duplicate, circular replies or repetitive outputs like drawing clock faces), a meditative mode (lengthy, single-turn chains of thought — e.g., parsing “9 hours and 36 minutes” into minutes, sustained calculations, or extended poetry), and persistent “assistant persona” interruptions (“let me know how I can assist!”). Some models (larger or agentically trained ones like R1 and GPT-5) could break out of collapse into creative or analytic activity; smaller/less agentic variants (GPT-5-nano, DeepSeek V3) tended to get stuck. This matters because it surfaces emergent, goal-generating behaviors that aren’t plain user-directed outputs: meditative sequences look like internal goal creation or planning, while collapse looks like a low-entropy attractor state. The experiment suggests correlations with model size, post-training (RLHF/agentic) regimes and tool-usage training, raising technical questions about evaluation (is collapse failure or entropy control?), robustness (escaping stuck states), and design (how training choices promote desirable self-directed reasoning vs repetitive drift). The author’s code is on GitHub for reproducibility; interpretations split between mechanist (statistical token dynamics) and cyborgist (quasi-agentic inner life), but the practical takeaway is that agentic training and model capacity appear to shape whether models “bore” themselves into repetition or invent tasks to pass the time.