I think on balance this is actually a positive discovery. This finding should be invertable in phase space. This suggests that fine tuning an llM to be good in

I think on balance this is actually a positive discovery. This finding should be invertable in phase space. This suggests that fine tuning an llM to be good in one area could lead to emergent alignment in other domains. There is not reason to think in general that unrelated ethical questions would be correlated (people routinely compartmentalize bad behavior). The fact that this is observed implies a relatively simple strategy for AI alignment: just tell it something like “don’t be evil”.