Rob Bensinger comments on Garrabrant and Shah on human modeling in AGI