Is it possible that GPT-6 could possibly be highly reliant on RAM? For example running smaller long context models on the CPU? Just a thought, I could be far off.