This is on GPT-J which has 6B parameters. I wonder if this scales well to much larger models like Llama 65B or GPT3