This argument for the proposition “AI doesn’t have an advantage over us at solving the alignment problem” doesn’t work for outer alignment—some goals are easier to measure than others, and agents that are lucky enough to have easy-to-measure goals can train AGIs more easily.
This argument for the proposition “AI doesn’t have an advantage over us at solving the alignment problem” doesn’t work for outer alignment—some goals are easier to measure than others, and agents that are lucky enough to have easy-to-measure goals can train AGIs more easily.