LLMs can handle much larger learning rates by letting networks naturally reduce their loss landscape sharpness