x
How hard is it to inoculate against misalignment generalization? — LessWrong