x
Metagaming matters for training, evaluation, and oversight — AI Alignment Forum