The competition in large models is fierce, with Llama-1 and 2k already outdated. The new generation of GPT-4 and Claude2.1 is leading the trend. The author discovered through extreme testing that most people did not use large models correctly and failed to unleash the AI's potential. The Claude team offers a simple solution: by adding one sentence, the performance jumped from 27% to 98%. Domestic large model companies also participated in the testing, with the Kimi team from Moon's Dark Side achieving good results using different methods.