x
Disincentivizing deception in mesa optimizers with Model Tampering — AI Alignment Forum