How to infer the method to OpenAI’s madness