[bigmadshoe]

You raise a good point. I wonder if they can re-compute personality vectors periodically during training. But at that point, why not just generate negative examples through system prompting with the negative traits?