Recently, domestic models have been using their computing power for training in the middle of the night, and their intelligence drops significantly at night—basically unusable. Now I use domestic models during the day and overseas LLMs at night, an AI version of time zone arbitrage.