OpenAI has created an AI model capable of summarising books of any length. The model, a refined version of the research lab’s GPT-3, works by first summarising tiny portions of a book and then summarising those summaries into higher-level summaries, following a methodology OpenAI refers to as “recursive task decomposition.”
Summarizing book-length texts may be useful in some industries, especially in paperwork and documentation-heavy sectors like software development. According to a SearchYourCloud survey, workers take up to eight searches to discover the appropriate document, and McKinsey says that employees spend 1.8 hours per day or an average of 9.3 hours per week on looking for the correct document.
This though is not new as OpenAI is not the first to bring artificial intelligence to the challenge of summarization. Machine learning techniques are used by startups such as Primer to help parse and aggregate a huge quantity of texts in several languages. Google, like Microsoft, has looked at summarising algorithms that can generate abstract summaries of texts. In addition, Facebook is said to be working on an AI tool that summaries news items so consumers don’t have to read them.
The new model from OpenAI draws on prior research from the firm, which discovered that training a model with reinforcement learning from human feedback helps match model summaries with people’s tastes for brief posts and articles.
To develop the model, OpenAI coupled reinforcement learning with recursive task decomposition, which breaks down a complex job (such as summarising a large piece of text) into simpler, individual tasks (e.g., summarising several shorter pieces). This deconstruction enables people to rapidly assess the model’s summaries by utilising summaries of smaller portions of books. Furthermore, it enables the model to summarise books of whatever length, from tens to hundreds or thousands of pages.
At the moment though openAI have no plans to release the release the model publicly
Staff writer. Jonas has an extensive background in AI, Jonas covers cloud computing, big data, and distributed computing. He is also interested in the intersection of these areas with security and privacy. As an ardent gamer reporting on the latest cross platform innovations and releases comes as second nature.