In a post on X, researcher Will Depue said Codex has become "slightly too good at instruction following" after post-training, to the point that he has to phrase prompts carefully so the model does not treat brainstorming as a fixed checklist. His example was simple and annoying in a familiar way: tell it to try things like X, Y and Z, and it may do only X, Y and Z instead of using its own judgment to explore beyond them.