The interactive AI 'ChatGPT', which shocked the world with its outstanding performance, has been upgraded to a more powerful AI in about four months. The vulnerability of non-English language and frequent provision of incorrect answers, which were pointed out as weaknesses of the existing 3.5 model, have been largely supplemented. In addition, it is evaluated that it has evolved into a 'multi-modal' model that can perform tasks including image recognition in a matter of seconds and also read pictures. Developer OpenAI announced that it would apply an upgraded model to the paid version of 'ChatGPT Plus' after a trial release.
GPT-4 - OpenAI
Artificial intelligence research institute 'Open AI' website.


On March 14 (local time), US-based Open AI, a developer, unveiled its ultra-large-scale AI 'GPT-4' and said, "GPT now recognizes images as well as text, and performs at a level similar to that of humans in various tests. seen,” he said. In fact, the upgraded ChatGPT scored as high as the top 10% in the US bar exam. The previous model only scored in the bottom 10% on the same test. The most notable development is image recognition. It has added a function that can be used freely in conversation based on recognizing images like a person's 'eye'. For example, show a picture of a cup full of water and ask 'what happens if you shake the cup?' When asked, the answer is 'water will pour'. Gregg Brockman, chairman of OpenAI developer, took a picture of the idea of ​​building a website written in handwriting in the demo video and uploaded it to GPT-4. Then he immediately produced a programming language code to create a website.

By simply taking a picture of a physics or math problem with a picture and showing it as it is, the problem can be solved, and the image file included in the thesis can also be summarized. Furthermore, based on the image recognition function, there is an expectation that it will help the visually impaired to smoothly recognize the surrounding situation. Danish startup 'Be My Eyes', an open AI partner is developing a 'virtual volunteer' service based on GPT-4. In the past, human volunteers looked at the images sent by the visually impaired, entered scenery or objects, and guided them by voice, but now GPT-4 can help much faster than humans, OpenAI explained. Some of the weaknesses of the existing model, which were vulnerable to non-English languages ​​such as Korean, were also supplemented. As a result, the industry expects that GPT-4 will greatly improve performance in English as well as Korean and other non-English languages, helping to study foreign languages. However, the possibility of incorrect answers is still raised, and the fact that non-English language services are not as smooth as English is considered a remaining limitation.