OpenAI Unveils Advanced AI 'GPT-4'
Top 10% Level in US Bar Exam
Multimodal Capabilities... Assistance for Visually Impaired
American artificial intelligence (AI) startup OpenAI has unveiled the super-large AI model 'GPT-4.' It has elevated expertise in specific fields to a human level. It has begun to understand not only text but also images. It is regarded as an impressive product that companies are willing to pay for to boost productivity.
On the 14th (local time), OpenAI released GPT-4. It is an advanced version of GPT-3.5, which was applied to the AI chatbot 'Chat GPT.' This comes just four months after the release of GPT-3.5 in November last year.
The most noticeable change is the capability for multimodality. Multimodal AI can understand and generate complex information such as text, images, videos, and audio. Unlike GPT-3.5, which communicated only through text, GPT-4 understands both text and images simultaneously. In a demonstration video, when shown an image of a squirrel taking a photo of an acorn with a camera, it described this in text and even added that it was 'funny' that the squirrel was holding the camera like a person. However, it cannot convert text into images. Another AI developed by OpenAI, 'DALL·E 2,' can create a drawing when given a prompt like 'a girl holding flowers.'
OpenAI introduced 'Be My Eyes' as the first partner of GPT-4. It is an 'AI volunteer' that verbally informs visually impaired people about what appears on their phone screens. For example, if a photo of the inside of a refrigerator is taken, it tells what is inside; if a video of how to use a machine at the gym is taken, it explains which buttons to press.
Compared to GPT-3.5, GPT-4 is smarter. GPT-4 passed the mock exam for the U.S. bar exam with a score in the top 10%, demonstrating the potential to compete with humans in specialized fields. It scored in the top 7% and 11% in the reading and math sections of the U.S. SAT, respectively. Its language abilities have also evolved. In a demonstration video, when given a long document and asked to summarize it with sentences starting with 'G,' GPT-3.5 failed, but GPT-4 completed it neatly. Ha Jung-woo, head of the AI Lab at Naver Cloud, said, "It has evolved into a very powerful tool for improving productivity." He added, "It has developed significantly in specialized knowledge areas, making companies willing to pay to use it."
GPT-4 can remember eight times more conversation than 3.5. While GPT-3.5 remembers about 4 to 5 pages of a book during a conversation, GPT-4 can remember the length of a 50-page short novel. This means GPT-4 recalls and refers to information mentioned 35 pages earlier during a conversation.
The number of parameters, which determine AI performance, was not disclosed. Experts believe it has grown beyond the 175 billion parameters of GPT-3.5 but has certainly not reached the 100 trillion parameters suggested by some. The 100 trillion parameter scale is equivalent to the human brain.
However, limitations remain. The 'hallucination' phenomenon, where incorrect answers are presented as correct, has decreased but still exists. In internal factuality evaluations, GPT-4 scored 40% higher than GPT-3.5. While users could induce 'jailbreaks'?making the AI break rules?in GPT-3.5, GPT-4 is harder to deceive. OpenAI emphasized, "The important thing is that it still cannot be fully trusted," and added, "Especially in high-risk situations, careful attention is required, such as human review, providing additional evidence, or avoiding high-risk use."
Its Korean language ability is also limited. Although the number of translatable languages has increased to 26, it differs from super-large AI models specialized in Korean. For example, when asked whether riding a motorcycle on the highway is allowed, both GPT-3.5 and GPT-4 give the same vague answer: 'Generally possible, but regulations may vary by country or region.' In contrast, Naver's 'HyperCLOVA X' responds, 'In Korea, motorcycles are prohibited on highways and roads designated for automobiles.'
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.


