AI Services as a Research Paradigm — LessWrong