Data has always been key to LLM success, but it's becoming key to inference-time performance as well.