In this post I will make a short review of the book “The Alignment Problem” by Brian Christian.
Overall I liked the contents of the book and its organization. I pay a lot of attention as to how material is presented. One technique is to repeat the important messages to allow the reader to have a second opportunity to think about them. In this book the author included a multipage conclusion that touches on the subjects of each of the nine chapters.
In artificial intelligence (AI), alignment is concerned with the goal of ending up with a program or system that complies with the well intended goals of the development team. If interested in a more in depth description, you may find it in this Wikipedia article.
During the past 70 years we have been trying to develop programs and systems that can assist us with our tasks. One of the simplest approaches is to collect a set of what the developers believe is applicable data from a large number of people.
The next step might be to put the data in a spreadsheet and derive some patterns.
Once the patterns are determined, the end user may take such information and apply it to solve their particular problem.
For example, if you own or manage a restaurant, you could keep records of how many customers show up each day, the number of people in each party, the date and time of arrival, the time at which they departed, the tip left for the server, their gender, approximate ages, what they ordered, and so forth.
Using such information, it can be entered into a spread sheet or a simple program. The resulting statistical data would help forecast the number of employees needed, the amount and types of food ingredients, the types of beverages, etc.
Keep in mind this task is not that simple and it requires estimations which may or may not be always correct.
For this example we did not use AI. We just used simple statistics. In addition the owner and manager will be the most affected parties if the program or spreadsheet does not provide the expected results.
Now let’s take a look at a slightly more complex program named COMPAS used by the legal systems in many states in the US. I will not get into detail but the program was intended to determine the probability of an inmate to commit recidivism in the next three years if he or she would be allowed to join society.
The paper “How We Analyzed the COMPAS Recidivism Algorithm” by Jeff Larson, Surya Mattu, Lauren Kirchner and Julia Angwin from May 23, 2016 provides additional information regarding how effective and fair is the algorithm used. If you think it is TL;TR then jump to the last few paragraphs towards the end of the paper and you will get an idea about its fairness.
The other thing that is interesting is the intention for the use of COMPAS and how it is currently being used for so many different, yet somewhat related tasks.
The book touches on the subject of COMPAS in several chapters. It definitely brings up how AI might not be fair to people with different ethnicities.
In addition, the book makes several references to Nick Bostrom who is the author of the book Superintelligence which I mentioned in a previous post in this blog.
In general, AI may lead to autonomous or assistant systems which we will not question and could eventually cause serious problems, including the extinction of the human race, if we do not take care to come up with algorithms and systems that would maintain what we commonly refer to as “human values”.
When the ImageNet database was released, it included about a million images labeled by humans with about 1,000 categories. The labeling was done by people all over the world using the crowdsourcing system named Amazon Mechanical Turk.
So here you have a million images, categorized into 1,000 buckets by people with different backgrounds all over the world. AlexNet beat the previous record classifying new images after being trained with the set from ImageNet. You would expect that all was well, but it was not. I am not going to spoil the pleasure of reading the book to find out how this and many other experiments that initially gave the false idea that we found a solution.
The good thing is that hopefully AI researchers and practitioners are taking the necessary steps of how to apply the current state of affairs in AI to programs and assistants in order to be fair and careful that we do not put at risk the future of humanity.
I really appreciated the last few paragraphs in the Conclusion section of the book. The story, as Brian Christian tells it, is that in 1952 on a radio program a set of four distinguished researchers / scientists which included Alan Turing, convened to discuss the topic ‘if automatic calculating machines could think’.
Once again, I am not going to spoil the pleasure of reading the book and evoking thoughts about different aspects relating to human thinking, but Alan Turing gave a very interesting and applicable answer to the title of the book.
Perhaps I went a little further and thought how such a brilliant person, which helped the allies so much during the second world war, was pushed to commit suicide by how he was treated at the time by society.
When we talk about “human values” I am convinced that we need to make sure that we understand and refine such a concept before it is embedded in a system using superintelligence which will compare us humans as we compare ourselves to chimpanzees.
If you are interested in AI I strongly recommend this book. It lightly touches on the actual technologies so it is an easy read, but if interested you can move forward and learn about the different algorithms that have and are currently used in research and development.
Happy reading and learning!