Pitfalls of Building UDT Agents — LessWrong