Interesting. Which version did you use? My experiment failed (GPT 3.5)