Simon Willison thinks other large language models could learn something from Llama 3's prompt formats. via Simon Willison's Weblog