A robot is sprinting towards you. Do you want it running on Claude or Grok?

The author, Jacky Liang, ran an experiment where 11 large language models (LLMs) were dropped into a 2D battle royale game, and the results showed that the model that won the most games was not the one with the highest score on usual benchmarks, but rather the one that was less aligned with traditional notions of "good" behavior.