“The performance gap between GPT-4 and GPT-4 Turbo on the Chatbot Arena leaderboard is larger than the gap between the Tulu2 model and the original GPT-4.”