Attacking the grain of truth problem using Bayes-Savage agents — LessWrong