Ambiguous out-of-distribution generalization on an algorithmic task — LessWrong