Sufficiently optimized agents appear coherent — LessWrong