The ChatGPT boom is affecting the world. Last week, the developer OpenAI announced the latest generation of computing GPT-4 to improve the previous generation of GPT-3 problems.

(Reuters)

[Reporter Wu Baixuan/Taipei Report] The ChatGPT boom has affected the world. The developer OpenAI announced the latest generation of computing GPT-4 again last week to improve the problems of the previous generation of GPT-3. According to the analysis of Taiwanese experts, the computing power of the 4th generation is stronger than that of the 3rd generation. I read more than 25,000 words, and I can interpret pictures, etc., but it is still not completely true to reality. I am only an interpreter who can speak multiple languages. If you want to ask questions about right and wrong, you cannot completely rely on it.

Developed ChatGPT chat robot for OpenAI, which can automatically generate text answers through asking questions. , to answer questions more accurately, and once again cast a shock bomb on the development of artificial intelligence (AI), the Taiwan Science and Technology Media Center recruited experts for analysis.

Please read on...

Cai Zonghan, a professor at the Department of Engineering and Engineering at Central University, said that GPT-4 learns a large amount of text and images, acquires higher creativity and imagination, can generate more interesting and unique text and image descriptions, and has the ability to interpret pictures, and learn from images Generate text descriptions, but it still fabricates non-factual answers, which may be misleading or negative, and more fact verification and checking mechanisms should be introduced.

Li Yizhi, an assistant professor at the School of Communication, National Chengchi University, described that Google is a librarian who can only help you find things, but ChatGPT is a super multilingual interpreter who loves to "translate" or "translate" for you, and is more willing to add , if you want to use ChatGPT well, you should provide open content. If it is "right or wrong", it is best not to rely entirely on it.

Cai Yanlong, an associate professor and dean of the Department of Applied Mathematics at Chengchi University, said that GPT-4 can read more than 25,000 characters at a time, which is better than GPT-3's 2048 characters, and long article abstracts are better, but the mode is still "use the previous word Predict the next word", as Microsoft CEO Nadella said, it is the user's responsibility to produce useful and correct things, but it may make everyone think, only the media, publishing houses, or famous people who insist on checking the text is trustworthy.

Zeng Yuanxian, Distinguished Professor of the Institute of Library and Information Science, Taiwan Normal University, believes that misleading and biased information has not been deleted from ChatGPT and GPT-4, and may still be guided by some kind of prompt, and there will still be out-of-the-box, reasoning errors, and various prejudices. Nor does it understand what happened after the training data, but GPT-4 is 40% more capable of responding to facts than its predecessor.