How much data from o1 would DeepSeek actually need to actually make any improvements with it? I also assume they'd have to ask a very specific pattern of questions, is this even possible without OpenAI figuring out what's going on