LocalLLaMA probes prompt convergence across LLMs

The interesting part here isn’t that models agree, it’s why they converge so often on the same culturally “plausible” completion. That makes the prompt a decent litmus test for model priors, but a weak test of determinism.

• –“Guess a number between 1 and 50” tends to surface a human-biased midpoint, not a magical universal constant
• –Favorite-planet prompts like “Saturn” lean on common internet associations, so multiple models collapse onto the same high-frequency trope
• –Temperature, system prompts, and safety layers can flip the result, so cross-model comparisons need fixed decoding settings
• –The real pattern theme is not truth, but stereotype, salience, and training-data overlap
• –If you want a stronger experiment, use prompts with low semantic priors and run repeated samples across identical decoding parameters