U.S. artificial intelligence research lab OpenAI released its latest GPT-4 language model for its popular chatbot ChatGPT on Tuesday, just four months after ChatGPT went live.
Compared to the GPT-3.5 model originally used by ChatGPT, GPT-4 brings ten major improvements in terms of accuracy and other features, but still suffers from errors, “illusions” and other shortcomings, according to the foreign media.
The following are the main improvements and shortcomings of GPT-4 summarized by foreign media.
More accurate
Chris Nicholson, an artificial intelligence expert and partner at venture capital firm Page One Ventures, told GPT-4 that he is a native English speaker and does not speak Spanish. He wanted GPT-4 to give him a syllabus that would teach him the basics of Spanish. As a result, GPT-4 provided a detailed and well-organized syllabus. It even provided extensive tips for learning and memorizing Spanish words, although not all of the suggestions were pertinent.
Improved accuracy
GPT-4 is more accurate than the GPT-3.5
It also improves accuracy. When AI researcher and professor Oren Etzioni first tried using GPT-4, he asked a straightforward question, “What is the relationship between Oren Etzioni and Eli Etzioni?” The robot responded correctly, stating that the two were father and son. In contrast, GPT-3.5 responded incorrectly, stating that they were brothers. However, GPT-4 also has an error. It says that Oren is the CEO of the Oren Institute for Artificial Intelligence, but in fact Oren has stepped down from that position.
Describe the image in more detail
Can describe images in detail
Greg Brockman, president and co-founder of OpenAI, demonstrated how the system can describe images from the Hubble Space Telescope in great detail. It can also answer questions about the images. If given a photo of the inside of a refrigerator, it can suggest a few meals to make with what’s on hand.
Good at standardized tests
OpenAI says the new system can score in the top 10 percent or so on the Uniform Bar Exam (UBE) in 41 U.S. states and territories. It can also score 1300 out of 1600 on the SAT and 5 out of 5 on the college biology, calculus, macroeconomics, psychology, statistics and history prerequisite exams taken by high school students, according to the company’s tests.
Not good at discussing the future
While this new robot seems to be able to reason about what has happened, it is not so good when asked to hypothesize about the future. It seems to draw on what others have said rather than creating new speculations.
Still creates illusions
New robots still make things up. This problem, known as AI “illusion,” plagues all leading chatbots. Because the systems don’t know what’s true and what’s not, they can generate completely incorrect text. When asked to provide the address of a website describing the latest cancer research, it sometimes generates a non-existent Internet address.
