First, she makes extracts from the sections, which she summarizes in the text.
The artificial intelligence model is based on GPT-3. OpenAI has no plans to make it available or release the source code.
The model is part of the company's ongoing research to develop artificial intelligence; the researchers said.
First, the model divides the book into sections and makes a squeeze of each, which is combined into several more squeezes until the final text is obtained.
As an example, the company cites several books - for example, "Alice in Wonderland", in which 26 thousand words, the model has reduced to 66 pomaces of 6 thousand words, then to six by 830 words. There are 136 words left in the final retelling.
OpenAI trained the model of art books in a GPT-3 training dataset, which averaged over 100,000 words, says Vivian Blumenthal
To evaluate the model, the researchers took the 40 most popular books of 2020 and assigned two people to read them and write a squeeze, and then evaluate each other and the model.
The model successfully produced "book-level" texts that contained most of the information, but sometimes generated inaccurate statements due to a lack of context, the researchers say.