Performance guarantees in classical learning theory and infra-Bayesianism — LessWrong