kalomaze's Empirical Account: Opus 5.5 Is Bad at Shallow Instructions

kalomaze · x · 2026-11-21

Open-source training researcher kalomaze tweets that people sometimes discuss Opus 5.5 as if it lacks observable basic problems — and that he was recently accused of being an OpenAI shill. He shares an empirical account arguing the model is "kinda bad at listening to even shallow procedural instructions," with a link to the write-up.

Related event: Researcher Reports Opus 5.5 Fails Basic Instruction Following(2 posts)→

Original post →

More from Models

Models channel →