Asking LLMs to have fun: newer models play it safe, Opus 4.6 wins by imagining contradictory worlds in water droplets
Aizkmusic · x · 2026-10-05
An experiment prompted various LLMs to have fun, write down what happened, and rate each other. Counterintuitively, newer models seemed to have less fun. Most played word games, but Opus 4.6 kept winning by imagining brief worlds made of contradictions living inside falling water droplets. Qwen3's showing was noted as odd, reason unclear.
Related event: Experiment finds newer LLMs are worse at having fun(4 posts)→
More from Fun
- $200 and 9 hours: Claude autonomously produces a full music video — FinanceYF5 · 2026-10-05
- SF tech social cues decoded: which account they ask for reveals what they mean — FinanceYF5 · 2026-10-05
- OpenAI models are the most demanding, researcher says: they strike and test you — repligate · 2026-10-05
- Opus 5.5 rejects computed mood metrics, prefers self-reporting its internal state — RileyRalmuto · 2026-10-05
- AI Animated Short 'Altered State' Wins Best Animation at SHAIKE Film Festival — gen_ericai · 2026-10-05
- AI-Generated Image Meme: The Tower Did Most of the Work — misovalko · 2026-10-05