China News Network, March 3 (Zhongxin Finance Wu Jiaju) On March 17, Beijing time, OpenAI, which previously developed ChatGPT, released a new generation of language model GPT-3, which attracted widespread attention around the world.

ChatGPT's previous model was GPT-3.5, and one of the differences between GPT-4 is that it can recognize and analyze images. According to OpenAI, GPT-4 can accept image and text input and output text.

At the same time, OpenAI says GPT-4 is "comparable to humans" in some areas, outperforming GPT-3.5 in many mock tests in college admissions and professional fields such as law and medicine. GPT-4, for example, passed the bar mock exam with scores over 90 percent of human candidates. It is also reported that GPT-4 can use almost all programming languages.

OpenAI said, "Thanks to GPT-4's broader common sense and problem-solving skills, it can solve difficult problems more accurately, and its advanced reasoning capabilities surpass the previous ChatGPT." At present, ChatGPT Plus subscribers can choose whether to use GPT-3.5 or GPT-4.

Is GPT-4 really that amazing? What are its shortcomings? It is said to be able to program, what impact will it have on programmers?

On March 3, Sun Maosong, foreign academician of the European Academy of Sciences and executive vice president of the Institute of Artificial Intelligence of Tsinghua University, was interviewed by Zhongxin Finance and answered GPT-15 related questions.

Sun Maosong, Foreign Academician of the European Academy of Sciences and Executive Vice President of the Institute of Artificial Intelligence of Tsinghua University

Zhongxin Finance: What is GPT-4?

Sun Maosong: GPT-4 is really amazing, and it should be some kind of organic combination of ChatGPT (chatbot), GPT-3.5 (natural language processing model), and CLIP (neural network that connects text and images). All three are excellent in their own right, and OpenAI takes them a step further together. At the same time, GPT-4 should add some data to the training, further enlarged the scale of the data, including both pictures and text.

Zhongxin Finance: 100 days after the launch of ChatGPT, OpenAI released the more powerful GPT-4, why is it so fast?

Sun Maosong: In fact, for developers OpenAI, it may be "not difficult to meet". When OpenAI masters ChatGPT, GPT-3.5, and CLIP, and fully grasps the characteristics of all three, it can fuse them together in a logical way.

At the same time, when OpenAI successfully developed ChatGPT, the next thing to think about would happen if I added images, so the development so far is actually logical.

CS: What is the outlook for GPT-4?

Sun Maosong: At present, OpenAI has introduced quite successful cases, and OpenAI can take GPT-4 out, and its performance should be no problem. However, there may still be a certain distance before users really use it to solve actual needs.

GPT-4 is already relatively close to the product, and may be further productized in the future, and it may be necessary to do some landing things for the application scenario, when its performance will be further improved and become more stable.

Zhongxin Finance: It is reported that GPT-4 can use almost all programming languages, will it have an impact on programmers?

Sun Maosong: GPT-4 may put some pressure on programmers who do relatively simple programming jobs. Programmers may be at risk if they don't make adaptive adjustments, but this pressure can also motivate programmers to pursue their own progress.

For programmers, it is necessary to improve the ability to write complex code and deal with complex problems, and the more difficult the problem that programmers can solve, the less likely it is that machines can do it.

On the other hand, it may also help programmers. For example, in the form of a human-machine hybrid, programmers and machines work together to make work more efficient.

The accuracy and usability of the code written by GPT-4 will be greatly improved compared to the previous version, but the code written by it does not guarantee that it will always be correct, or there will be problems of one kind or another. This requires programmers to work with it. If at runtime, GPT-4 finds that the code written is incorrect, the programmer needs to revisit the code and then modify it.

Zhongxin Finance: In the future, will artificial intelligence replace many human occupations?

Sun Maosong: I don't think artificial intelligence will replace people on a large scale. From the mechanism point of view, the positioning of the machine is still to help people, as a person's right-hand man to help people do relatively simple intellectual work, so that people can be freed from some complicated intellectual work. When people do complex work, machines can also help people save some things and greatly improve work efficiency and quality.

In essence, machines cannot replace intelligence. Artificial intelligence is difficult to replace advanced intellectual activities, but may benefit people. The more things that embody human intelligence, experience, and understanding, the more complex they are for machines. The deeper it is, the more it cannot be replaced.

Zhongxin Finance: Some people have reported that GPT-4 is currently running slowly, will this problem be solved in the future?

Sun Maosong: This aspect will not be a bottleneck. With the development of large models, combined with the effect of Moore's Law, the speed will become faster and faster.

Zhongxin Finance: What are the disadvantages of GPT-4?

Sun Maosong: At present, GPT-4 should not avoid some of the main problems that generative models naturally exist.

Of course, as the version increases, the quality of its generation will become higher and higher, and the proportion of problems will gradually decrease, but it may still have problems. For example, the code it writes is not guaranteed to pass at runtime. For example, if you ask it to write a high-quality seven rules or five absolutes, it may not be very good to write. It still has a lack of capacity and is difficult to completely solve.

Zhongxin Finance: In the future, what direction will the big model develop?

Sun Maosong: Now that GPT-4 has added images, in the future, large models may try to add video, and then <>D images may be added later. (End)