Goldilocks and the Three Optimisers — LessWrong