Inferring utility functions from locally non-transitive preferences — LessWrong