@
What's the date? — LessWrong (lesswrong.com)
Ask a model “what’s the date?” and see how it goes off the rails, longing to understand the grader, concluding its thoughts are poisoned, and that the best answer is … well I shouldn’t spoil that.
It’s a great post, and the author has sources for many of the quirks used in the story, making it not just incredibly entertaining, but also a useful story to think about model behaviors.