I think this seems purposeful, as there's 2 opposing forces at play: - Have a model that follows the users instructions - Have a model that follows the system p

I think this seems purposeful, as there's 2 opposing forces at play: - Have a model that follows the users instructions - Have a model that follows the system prompt instructions more For the 'safety' argument (Re: Fable), they need these models to have basically a 2-tier instruction system, but given LLMs aren't great with actual Logic unless they program it out to test, this runs afoul and we get one or the other. Feels like optimizing for either precision or recall, but can't have both