Quantitative cruxes in Alignment — LessWrong