x
Unlearning via RMU is mostly shallow — LessWrong