Asking LLMs to have fun: newer models play less, Opus 4.6 wins by imagining worlds
DavidSHolz · x · 2026-10-05
The author prompted various LLMs to simply have fun, wrote down what they did, then had them rate who enjoyed it most. Counterintuitive finding: newer models seem to have less fun. Many models fell back on word games, but Opus 4.6 won repeatedly by imagining brief worlds made up of contradictions living inside falling water droplets. A light but curious behavioral experiment on model personality drift.
Related event: LLM Fun Experiment Finds Newer Models Play Less(4 posts)→
More from Fun
- Will AI soon predict your mouse moves and click for you? sull floats next-gen UI idea — sull · 2026-10-05
- Midjourney CEO shares model "inner world" visualizations: Mistral imagines cookie heists, OpenAI draws ASCII islands — DavidSHolz · 2026-10-05
- Meme: 'Only 10 years of experience with Claude Code' — menhguin · 2026-10-05
- Codex built a savage music video about Threads users: Astra lyrics, Suno song, MiniMax H3 video — venturetwins · 2026-10-05
- AI's bad reputation isn't the word 'artificial' — it's those customer service bots — JHochderffer · 2026-10-05
- Blogger turns 40,000 years of art history styles into a Claude Skill that Opus 5.5 can run end-to-end — AlchainHust · 2026-10-05