Back to browse
LLMs' Favorite Colors

LLMs' Favorite Colors

by gimlids·Apr 6, 2026·2 points·0 comments

AI Analysis

●●SolidRabbit HoleNiche Gem

Sampling hex codes from 10 LLMs exposes bias patterns useful for fingerprinting.

Strengths
  • Concrete data: 200 samples per model across 10 different open-weight LLMs.
  • Visual fingerprinting hypothesis offers a novel way to detect model distillation.
  • The simple treemap visualization makes frequency distribution across models immediately readable.
Weaknesses
  • Fingerprinting claim remains speculative without further validation or established false-positive rates.
  • Limited utility beyond curiosity; no API or tool to replicate analysis.
Category
Target Audience

AI researchers, LLM enthusiasts, curious developers

Post Description

I sampled CSS hex colors out of 10 open weight LLMs to build each model's favorite colors palette, or its flag... Interesting results?

Similar Projects

AI/MLMid

My "home rig" for iterative attribute-weighted LLM benchmarking

Home rig for attribute-weighted benchmarking lacks the polish of established eval frameworks.

Ship It
yuvalhaim
211mo ago