My understanding of the core alignment problem is giving an AGI any goal at all (hence the diamond-alignment problem). A superintelligent AGI will know better than we do what we desire, so if we simply had the ability to give the AI instructions in natural language and have it execute them to the best of its ability, we would not have to figure out the correct human values.
My understanding of the core alignment problem is giving an AGI any goal at all (hence the diamond-alignment problem). A superintelligent AGI will know better than we do what we desire, so if we simply had the ability to give the AI instructions in natural language and have it execute them to the best of its ability, we would not have to figure out the correct human values.