[AN #66]: Decomposing robustness into capability robustness and alignment robustness — LessWrong