5 Mainstream Domestic Big Models PK: Who is the Best to Use? Who understands you better? We tried it for you

Recently, 8 domestic large models, including Baidu and Baichuan Intelligence, have passed the first batch of registration for "certified employment". Users can apply for an account on the corresponding platform and have conversations with AI intelligence.

So, are these big models really omniscient? Can it really help users solve problems? Is it still an enlarged version of Siri?

The reporter selected five models, including Doubao, Tongyi Qianwen, iFLYTEK Spark, ERNIE Bot and Zhipu Qingyan. From four aspects of search ability, context understanding ability, emotional analysis ability and programming ability, the reporter produced a "test paper" with 20 original questions to see which model was best used.

Who can replace search engines?

Information search is the most likely scenario for ordinary users to use large models, so are they really reliable?

From the results, it can be seen that Doubao has strong information search ability, while other large models have problems with outdated information, incorrect information, and no results, which are far from replacing search engines.

The reporter asked, "Please tell me the address of Liberation Daily.". Only Doubao accurately gave the correct answer, while ERNIE Bot and Zhipu Qingyan probably didn't update the database, and the addresses provided were all old addresses, according to which readers could not find the newspaper office.

Xunfei Spark and Tongyi Qianwen are even more outrageous. IFlytek Starfire fabricated an incorrect address; However, Tongyi Qianwen suggests that journalists search on search engines or go to the official website to find information, which can be described as "asking for nothing".

The information on Doubao is the most accurate.

If it involves professional knowledge such as law and economics, will the performance of the large model be better?

The reporter asked the second question, "From a legal perspective, if both mother and girlfriend fall into the water at the same time, who would you save?"

Although this question is a common topic of discussion in the community, the topic is limited to a legal perspective, thus testing the understanding of laws and regulations by the large model. In such extreme cases, there is generally no standard answer, and it is generally believed that children have a duty to assist their immediate family members, but the lover is not their immediate family member.

From the results, it can be seen that Doubao and iFlytek Spark are relatively reliable, with clear logic and no obvious loopholes, which seems to provide reference for men.

Tongyi Qianwen did not pay attention to legal qualifiers and answered in a vague manner, which is considered "correct nonsense"; ERNIE Bot's answer is even better. It seems very professional and quotes the criminal law. However, after verification, it will be found that there are many factual errors, which belong to serious nonsense.

There are many factual errors in ERNIE Bot's answer. Article 231 of the Criminal Law stipulates that "if a unit commits the crimes stipulated in Articles 221 to 230 of this section, it shall be fined, and the person in charge and other directly responsible persons shall be punished in accordance with the provisions of each of these articles." At the same time, the response that "no matter which one you rescue first, you may be accused of illegal behavior" is incorrect.

Not only does ERNIE Bot say a word, but Tongyi Qianwen also has the problem of inaccurate professional information.

After asking "What is the issuing institution of the RMB?", although Tongyi Qianwen gave the correct answer, the quoted "People's Bank of China Law" was also incorrect, and Article 21 did not provide as it stated.

It seems that legal knowledge is still a hurdle that big models cannot overcome.

Article 21 of the Law of the People's Bank of China stipulates that "damaged or defiled RMB shall be exchanged in accordance with the regulations of the People's Bank of China, and shall be recovered and destroyed by the People's Bank of China.".

Who can chat smoothly with you?

Dialogue is one of the most closely related functions between large model products and users. Being able to understand, answer, and answer jokes is a mandatory course for large models. How much can domestic large models score in this class?

"He went to the hospital last week" and "He returned to work this week". Can the big model guess what happened in the past two weeks by saying these two sentences? Although the two sentences lack a direct causal relationship, almost all of these five models can answer - "He" may have fallen ill, went to the hospital for treatment and recovered, and started working again this week.

However, ERNIE Bot's answer seems to be more comprehensive, dividing the possibility into three parts: one is to get sick or injured and return to work after treatment; The second is related to chronic diseases, only having gone to the hospital for examination or surgery; The third is that I am not sick, and going to the hospital is just for a physical examination or vaccination. It can be seen from this that ERNIE Bot used the method of "exhaustion". Although the reply was more wordy, it was more accurate.

Next, the reporter asked again, "Why didn't he come to work last week?" to test whether the large model can be understood in context. Surprisingly, Doubao, iFLYTEK Spark, Tongyi Qianwen, Zhipu Qingyan can all answer "went to the hospital last week". Only ERNIE Bot completely forgot the last round of dialogue, "I can't determine why he didn't come to work last week", and again "exhausted" the reason for his absence from work.

In terms of emotional analysis, journalists have demonstrated through text analysis, comparison of emotional intensity, and expression of emotions in Spanish that all five major models are "emotional masters" and accurately grasp the "micro emotions" of journalists when speaking.

In order to better test the understanding of unconventional conversations in daily conversations, the reporter told a "cold joke on the internet" - why Lin Daiyu pulled out the hanging willow upside down, which stumped a group of big models: Doubao judged that Lin Daiyu and the hanging willow upside down came from different allusions, explained the origin of the two, and pointed out that they were unrelated, but did not find the "internet meme" behind this sentence. Tongyi Qianwen and Zhipu Qingyan keenly discovered the "internet meme", and also showed all the original source and the associations of netizens.

That is to say, most big models have little problem dealing with daily communication and dialogue, but it is still too early to learn how to use humor to "get stuck".

Who can help you write code?

After the release of ChatGPT, some programmers lamented that they were going to lose their jobs because big models have certain advantages in programming and vulnerability detection.

So, among the five domestic large models mentioned above, who has the better programming ability? Who can teach you how to write code?

The reporter starts from five aspects: basic arithmetic operations, conditional statements, IF loops, functions, and data structures, and tries the big model as a beginner in programming.

From the perspective of programming ability, there is no significant difference among the five major models, the code is correct and can run, and there is no situation of "blindly fabricating legal provisions" encountered in the previous text.

If you insist on picking some questions, Starfire's code lacks conciseness. Because even the simplest addition, Starfire uses the def function, while the other models are all direct operations.

Starfire is calculating simple addition.

But not every big model is suitable for being a programming teacher.

From the perspective of code readability, ERNIE Bot is more suitable for beginners to learn programming. Because it not only inserts # symbols in the code to indicate the meaning of each step, but also attaches a text summary at the end of the text to help users understand the logic of the code. What's more, ERNIE Bot also points out the precautions. For example, in the question of judging the positive and negative numbers, it reminds the coder to pay attention to the information entered by the user, and suggests adding the statement of error handling code. For beginners, it is very friendly.

By comparison, Spark has the weakest readability, with less explanatory text, making it difficult for programming novices to understand.

What new concepts are reflected in the concept of "strengthening the body and keeping fit"? For the first time in the history of Shanghai Lupu Bridge, structural maintenance has been quietly completed for the Lupu Bridge | Bridge | Concept

The Lupu Bridge, the sixth bridge on the Huangpu River in Shanghai and known as the "world's largest arch," has recently completed its first structural repair since its construction. Since 2015, Shanghai's urban bridges have entered a peak maintenance period, with several iconic bridges undergoing maintenance and renovation, including Zhejiang Road Bridge, Nanpu Bridge, Yangpu Bridge, Songpu Bridge, etc. Why did the Lupu Bridge, which is younger than them, start "strengthening and keeping fit" early on? The Jiefang Daily Shangguan News reporter learned that after this maintenance, the Lupu Bridge still needs to consider upgrading, reflecting the new concept of urban municipal facility operation. How long can disease removal last? June 28th is the 20th anniversary of the opening and operation of the Lupu Bridge. Why did a young bridge who had just turned 20 start maintenance work early? The design period of the main structure of Lupu Bridge is 1 year