Perils of under- vs over-sculpting AGI desires — AI Alignment Forum