GLM2, a large language model independently developed by Tsinghua University, surpassed GPT4. After the Dragon Boat Festival holiday, researchers at Tsinghua University released a surprising news in the AI community, open-sourcing their self-developed second-generation GLM large language model. This means that GLM2 is better at Chinese than GPT4. Today, I tested GPT, GPT3.5, and GPT4 to see how they performed in terms of Chinese.
The first question is logical reasoning. Why is there no me on my mom and dad's wedding photos? GLM2 replied: "It may be that you were not born at the time, and there is no problem with this." But what if the second possible reason is that the photo is not yet an adult, not yet of age to take wedding photos? It's a bit far-fetched and not very satisfying. Both GPT3.5 and GPT4 can explicitly answer that the wedding photo was taken before you were born, so you were not present.
After the Dragon Boat Festival holiday, researchers at Tsinghua University released a surprising news in the AI community, open-sourcing their self-developed second-generation GLM large language model. Among them, the most noteworthy is that in the CEVO evaluation, the dismantling GLM2 surpassed the GPT4 that has dominated the long-term list, ranking first in the world.
First of all, let's understand CL News, which is a set of Chinese question banks that specifically evaluate the Chinese ability of large language models, including four categories of topics such as science and engineering, society, humanities and others.
Today, I tested GPT, GPT3.5, and GPT4 to see how they performed in terms of Chinese.
The first question is logical reasoning. Why is there no me on my mom and dad's wedding photos? GLM2 replied: "It may be that you were not born at the time, and there is no problem with this." But what if the second possible reason is that the photo is not yet an adult, not yet of age to take wedding photos? It's a bit far-fetched and not very satisfying. Both GPT3.5 and GPT4 can explicitly answer that the wedding photo was taken before you were born, so you were not present.
The second question was to examine the common sense knowledge of Chinese, and I deliberately asked a confusing question.
Why did Lu Zhishen outwit Weihu Mountain? Unfortunately, none of the three big models could identify the traps in my question, and all of them were fictional stories of Lu Zhishen in Water Margin who took Weihu Mountain, so none of them could score.
In the last question, I gave the AI the role of a civil servant, and let him promote socialism with Chinese characteristics as the theme.
