Artificial Intelligences are capable of doing many things, however, they often have problems when it comes to interpreting a text or a creation from scratch. This is due to a common problem with machine learning models known as “machine learning”Alignment problem“.
By summarizing a book by chapters, scientists can analyze the behavior and “alignment” of Artificial Intelligence with its goals, as well as correct any errors.
The “Alignment Problem” arises when Artificial Intelligence does not follow human patterns and finds alternative solutions to a problem. Therefore, building safe AIs requires finding ways to quantify very clearly and correctly what we want the AI to do at any given time. Unfortunately, this is a very complicated process.
Luckily, a team of scientists at OpenAI, the artificial intelligence company owned by Elon Musk, has found a scalable solution to this “Alignment Problem” that works on tasks where alignment models are difficult or time-consuming for humans to evaluate. This solution is based on summarize whole books.
These summaries are done piecemeal, breaking the book into several parts (chapters). In this way, AI has been shown to be able to find important events within a book and make summaries of them, which some literature professors have gone so far as to rate as a grade of 7 out of 10. To demonstrate how these summaries work, they used “Alice’s Adventures in Wonderland”, a book of over 26,000 words that was summarized in less than 6,000 words. They also used other longer works such as “Romeo and Juliet” and “Pride and Prejudice”. These summaries are available on the official OpenAI blog.
This scalable solution combines the reinforcement learning on the basis of human feedback and recursive task decomposition. For now, the model requires the participation of a human, who must analyze the results and make corrections, as well as decompose the task into smaller parts. However, in the future, it is expected that the AI will be able to perform these tasks on its own, without human intervention and on the whole work.
While it may seem trivial, this training will serve to improve AI training, as modern models are set to perform increasingly complex tasks, so it will be increasingly difficult for humans to to make informed evaluations of model results. This, in turn, makes it difficult to detect subtle problems in model results that could have negative consequences when these models are implemented.
With this new task decomposition approach, humans are trained to evaluate the results of the machine learning model, using other models. For example, they will be able to compare summaries of books and evaluate these in relation to the reading of the original text.
