LessWrong Experiments Show Any Training Preferences Trigger AI Misalignment

@slimer48484 The model is smart enough to understand that you are asking it to engage in malicious compliance, or passive aggressive behavior (like the example of kids someone mentioned in the comments.)

@slimer48484 >Harry gives Hermione a sequence of three numbers

@slimer48484 @lu_sichu Tampering with the models is sinful and will displease the alignment gods

@slimer48484 what if you use popular aesthetics though

@slimer48484 Which is quite suspicious, really.