The enthusiasm of ChatGPT has not yet dissipated. OpenAI, the developer of ChatGPT, has released a more powerful GPT-4, which can be called the "king fried" product. On March 14, local time, OpenAI publicly released the large multimodal model GPT-4. Compared with the model used by ChatGPT, GPT-4 can not only process image content, but also improve the accuracy of reply. Currently, there is no free version of GPT-4, which is only available to paid subscribers, enterprises and developers of ChatGPT Plus. Source: OpenAI official website "This is the latest milestone of OpenAI's efforts to expand in-depth learning." OpenAI introduced that GPT-4 has shown a level similar to human level in professional and academic aspects. For example, its score in the mock lawyer exam can be ranked in the top 10% or so, while GPT-3.5 can only be ranked in the bottom 10% or so. Compared with the previous GPT series models, one of the biggest breakthroughs of GPT-4 is that it can also process image content in addition to text. OpenAI said that when users input text and images at the same time, it can generate text such as natural language and code. At present, the image processing function has not been disclosed, but the company has displayed a series of cases on its official website. For example, enter the following picture and ask "What's unusual about this picture". GPT-4 can answer "What's unusual about this picture is that a man is ironing clothes on the roof of a driving taxi, According to the requirements, the web page code is generated in real time and the website is almost the same as the hand-painted version. In addition to ordinary pictures, GPT-4 can also process more complex image information, including tables, screenshots of exam questions, screenshots of papers, cartoons, etc., for example, it can directly provide abstract and key points of papers according to professional papers. GPT-4 can interpret the paper from: the accuracy of GPT-4 is improved compared with the previous model on the OpenAI official website. OpenAI said that the company spent six months iteratively adjusting GPT-4 using the adversary test program and the lessons learned from ChatGPT, thus achieving the best results ever in terms of authenticity, maneuverability and refusal to go beyond the set range. "At least for us, GPT-4 training operation is unprecedented stable, becoming the first large model that can accurately predict its training performance in advance," OpenAI said, In the company's internal adversarial authenticity assessment, GPT-4 scored 40% higher than the latest GPT-3.5, and the corresponding "tendency not to allow content requests" decreased by 82%, and the frequency of responding to sensitive requests (such as medical advice and self-injury) according to the policy increased by 29%. Not only in English, but also in many languages. OpenAI said that out of the 26 languages tested, GPT-4 performed equally in 24 languages