Modeling the impact of AI safety field-building programs — LessWrong