The OpenAI company introduced GPT-4, a new generative model of artificial intelligence, which lies in the basis of both well-known services such as ChatGPT and the Bing search engine, as well as new programs. According to the developers, the new language model will be “the world’s first experience of working with a high-performance artificial intelligence system.”
In its blog, OpenAI notes that the difference between GPT-4 and its predecessor GPT-3.5 will be barely noticeable in casual conversation, but the differences between the systems will become apparent when compared in more complex tasks. GPT-4 passes exams and various tests much better than its predecessor.
OpenAI notes that the system can understand both text and graphical information. For example, if you load a picture of balloons and ask what would happen if you cut their strings, it answers that they would fly away. OpenAI claims that the ability to simultaneously analyze text and images allows for the interpretation of more complex input data.
According to the developers, GPT-4 is also more creative and collaborative. A model can generate, edit text, and perform technical (such as writing code) and creative tasks, such as composing songs, writing scripts or learning a user’s writing style.
Compared to ChatGPT, which can generate up to 3000 words in response, GPT-4 will be able to generate up to 25,000 words.
In addition, the new model has more training data from more sciences. OpenAI notes that GPT-4 can be a personal tutor in many subjects, which will always pay attention to the student and will never lose their temper if the student does not understand something.
OpenAI said it is already working with a number of companies to integrate GPT-4 into their products, including Duolingo, Stripe, and Khan Academy. The new model will also be available on ChatGPT Plus (the paid version of the chatbot) and as an API.
The developers say they spent 6 months making GPT-4 safer and more consistent. According to OpenAI’s internal estimates, GPT-4 is 82% less likely to respond to requests for prohibited content and 40% more likely to provide actual responses than GPT-3.5.