Many tech giants are now competing to develop an all-purpose AI that can solve any problem and solve any task. Referred to as the alignment problem, they want to develop a model that has room to solve any problem you give it, no matter how challenging or time-consuming it may be.

Following the same line of thought, Elon Musk’s ‘capped’ for-profit company, Open AI, has developed an AI model which can summarize books of any length. It’s a fine-tuned model that uses natural language processing through GPT-3 and gives you the gist of a book of any length.

The model follows a paradigm that Open AI has named ‘Recursive task decomposition’. It first summarizes small sections of a book and then adds those summaries to create a more refined and precise summary. In a time when employees spend about 1.8 hours a day or 9.3 hours a week, on average searching for and gathering work-related information, this formula can prove a boon to productivity.

According to the Open AI team, the model “achieves a 6/7 rating from humans who have read the book 5% of the time and a 5/7 rating 15% of the time.”

“Open AI believes that this is an effective ‘recipe’ that can be used to help humans supervise many other tasks. A scalable solution to the alignment problem needs to work on tasks that are difficult or time-consuming for humans to evaluate”, said a spokesperson.

To create the model, the Open AI team combined recursive task decomposition with reinforcement learning, an algorithm that entails training a system to perform a task by rewarding desired behaviors and punishing the undesired ones. It then trained the model on a subset of books that were mostly fiction and contained over 100,000 words.

For better evaluation, the team took the 40 most popular books and assigned a couple of people to read each book, write a summary, and rate the summaries from the model and each other. According to the OpenAI team, the model “achieves a 6/7 rating from humans who have read the book 5% of the time and a 5/7 rating 15% of the time.”

However, there were certain limitations to it:

·      The model’s book summaries lack coherence - While the model successfully generated ‘book-level’ summaries containing much of the important information, the summaries often read more like a list of events from the book rather than a coherent summary.

·      Task decomposition could be fundamentally limiting - Task decomposition assumes that separate parts of the task can be completed independently. However, this may not be true for summarizing books where earlier details are only revealed later, as in the case of mystery books. Also, the model sometimes generated inaccurate statements due to a lack of context.

·      Training on higher height tasks may be difficult - In general, policy errors at lower levels compound at each composition task, ultimately leading to large errors on the top-level task. As the model’s curriculum and node sampling strategies were chosen in an ad hoc way, it may also be making training significantly more difficult, and curriculum choice may matter a lot as a result.

“This work is part of our ongoing research into aligning advanced AI systems, which is key to our mission,” Open AI researchers wrote in a blog post. “Our progress on book summarization is the first large-scale empirical work on scaling alignment techniques. Going forward, we are researching better ways to assist humans in evaluating model behavior, with the goal of finding techniques that scale to aligning artificial general intelligence.”

It would be interesting to see whether AI can replace this aspect of human intelligence too or get subsided by it.

Image by oxi-roxy from Pixabay