The all-knowing ChatGPT-4 actually lost to a junior high school student? !

ChatGPT-4 got off to a bad start

ChatGPT has been popular for more than half a year. In many people's minds, ChatGPT, who knows everything about astronomy and geography, is already a god-like existence. Many people even fear his abilities and worry whether the arrival of AI will have a negative impact on humans. But is this really the case?

Recently, a piece of news has made some people suspicious of ChatGPT again. Wei Dongyi, an assistant professor at Peking University known as "Wei Shen" and a gold medalist in the 49th and 50th International Mathematical Olympiad, asked a math problem, which attracted many scholars and math enthusiasts to try. The recent "hottest" ChatGPT-4 also tried to solve this problem, but ChatGPT-4 was unable to get the correct answer and was stuck in the middle of the steps many times and could not continue the calculation. However, in sharp contrast, among the three answers selected by the questioner Wei Dongyi as "better examples of solutions", one of them actually came from a student in the second grade of Shenzhen Middle School!

"ChatGPT-4 is no match for a junior high school student!" "A junior high school student solved a Peking University math problem that ChatGPT-4 got wrong?" For a while, everyone was attracted by various reports. Is the ChatGPT myth going to be broken? Is it just a marketing cover?

Why is ChatGPT so awesome?

Unlike intelligent voice assistants such as Siri, ChatGPT-4 does not directly use search engines to retrieve questions and select answers from the retrieved answers. ChatGPT is a language model that uses deep learning technology and neural network models to learn language patterns in order to generate and understand natural language.

Specifically, ChatGPT uses a neural network based on the Transformer architecture, and learns knowledge about the grammar, semantics, logic, and common sense of natural language by pre-training on a large amount of text data. When processing natural language tasks, ChatGPT encodes the input text into a sequence of numbers and encodes it through a multi-layer Transformer encoder. In the encoder, neurons at each level process different features of the input text, such as vocabulary, grammar, context, etc., in order to generate a vector representing the meaning of the input text. After multiple layers of encoding, ChatGPT will obtain a text representation vector, which is then converted into a probability distribution for generating text using a decoder to generate the corresponding output text. Through continuous pre-training and fine-tuning, ChatGPT can complete language tasks such as text generation, machine translation, and dialogue systems.

ChatGPT also has its inherent flaws

ChatGPT-4's language processing ability is based on previously input text data and is optimized through a large amount of data training, which means that it cannot answer some questions it has never seen before. For example, the math problem mentioned above, the solution is not regular. For ChatGPT-4, which has not yet been exposed to relevant information, without the support of previous data, it cannot complete this irregular operation, which explains why ChatGPT-4 cannot solve the problem.

In addition to complex math problems, ChatGPT-4 is also unable to answer some highly professional fields, such as the medical and legal industries. Doctors with rich clinical experience do not diagnose patients based on what is described in books, and experienced lawyers do not simply prepare files based on the code. These fields require long-term experience accumulation and human emotions to make plans. ChatGPT-4 cannot accept big data training, and AI is emotionless and cannot give a more humane answer.

In addition, the prompt of ChatGPT-4 reads: limited knowledge of the world and events after 2021, which means that ChatGPT-4 cannot answer questions with today's thinking. This is also an obvious disadvantage of ChatGPT-4. This disadvantage can only be solved after it is connected to the search engine and can respond to real-time content. For example, the new Bing chat that Microsoft is developing will be more based on the nature of the search engine, combined with the advantages of ChatGPT, to create an AI that can obtain real-time information.

ChatGPT should be used as a tool

ChatGPT does not have legal status. Although some countries have begun to discuss the establishment of some legal frameworks in the field of AI, AI is still a tool and cannot bear legal responsibility independently. This means that once the information provided by AI is wrong, the recipient of the information makes a wrong decision because of the wrong information provided by AI, and the consequences will not be borne by AI.

Therefore, AI at this stage is here to help us. We can make good use of it, but we will not be replaced by AI. When talking about his views on ChatGPT, Jack Ma said that technologies like ChatGPT have brought challenges, but technologies like ChatGPT are just the beginning of the AI ​​era. We should use artificial intelligence to solve problems instead of being controlled by artificial intelligence. Although human physical and mental strength cannot compare with machines, machines only have "cores" while humans have "hearts". The industrial age is knowledge-driven and a competition of knowledge; the digital age is wisdom-driven and a competition of creativity and imagination, leadership, responsibility, and independent thinking.

Although ChatGPT-4 cannot solve Peking University math problems, it is unquestionably the strongest AI interaction at present. It is just limited by certain reasons, and there are still many things that AI cannot do. As for the future development of AI, we don't need to panic, but should make good use of its development to better help us work and live.