Post BAGtNYsVxAujoN6ejw by carnage4life@mas.to
 (DIR) More posts by carnage4life@mas.to
 (DIR) Post #BAGtNYsVxAujoN6ejw by carnage4life@mas.to
       0 likes, 0 repeats
       
       I think the focus on alignment at frontier labs is misguided.If you don’t want Claude deleting emails, don’t give it that permission or require human approval.Trying to “teach” an unreliable LLM that deleting emails without asking is bad is a fool’s errand. This applies to all dangerous actions.
       
 (DIR) Post #BAGtNaQwAg2kdQDf28 by carnage4life@mas.to
       0 likes, 0 repeats
       
       The tradeoff is this contradicts the goal of replacing human workers with AI, if humans have to review every consequential decision.So alignment research is a band aid given the business goals of AI labs.