It is very hard to start reasoning what could have happened in those models but I am pretty sure that it was not intentional sabotage.
No kidding? What’s the state of the art for lists of these kinds of prompt biased performance? Or should I ask Opus to game it out with Muse and <Tencent>
Got it, I’m curious about other biases and prompt context results. First that come to mind are Gemini’s very different responses depending on gender in the question.
If it exists, there are others?