Why Public LLM Snapshots Mislead: The Gemini 2.0 Flash and Vectara HHEM Case

https://direct-wiki.win/index.php/When_a_Hospital%27s_Triage_Assistant_Gave_Dangerous_Advice:_Dr._Lin%27s_Story

When published scores stop matching reality: a concrete problem Many teams rely on vendor snapshots and third-party score tables to choose models. That worked in the internet age for CPU benchmarks, but not for large language models