Preventing model exfiltration with upload limits — LessWrong