Charlie Griffin comments on Can models gradient hack SFT elicitation?