The Best Book I've Read About AI Isn't in English
Last year I read a great book about AI. It was written as an introduction for a general audience, but filled with references to classic papers and enough details that I learned new things reading it. Specific problems and methods in actual practice were introduced in a straightforward way, with their strengths and weaknesses clearly laid out by example. I'd like to recommend it to all my friends, but there's just one catch - it's in Japanese.
The book in question is 働きたくないイタチと言葉が分かるロボット, or "The Lazy Weasels and the Robot that Understood Language", by Ai Kawazoe with illustrations by Ayumi Hanamatsu. The book follows a regular structure: each chapter begins with the ongoing story of the lazy weasels who think they can use robots to do their work for them, and every robot starts off promising but ultimately has limitations that make it imperfect and leave the weasels unsatisfied. The second half of each chapter steps back from the fiction to explain the technical details of the systems described, including plentiful citations. The fictional first half is written in a simple, whimsical style suitable for a children's bedtime story, while the second half's detail makes it more like an approachable college textbook. The balance between the two halves is, frankly, perfect.
To give an example, at one point the weasels hear from a chameleon that the ants have created a robot that can use knowledge about the world to answer any question correctly. The weasels test it by asking it how tall Mt. Fuji is, and it immediately responds accurately with "3776 meters". They go on to ask a variety of questions and are impressed by the comprehensive knowledge of the giant ant robot as it seems able to cope with whatever they throw at it.
At that point a lone tanuki wanders by and interrupts to ask the robot a question: "If Ms. Fukuko the Owl picks up a newspaper in the living room, walks through the kitchen and into the workroom, where is the newspaper?"
The weasels ask this weird tanuki who he thinks he is and why he's wasting the robot's time asking such a simple question, and then the robot answers: "the dining room."
This short exchange shows the promise and current limitations of the field known as Question Answering, best known through its application in IBM's Watson on Jeopardy. While powerful language models can solve some kinds of questions without any specially prepared data, the ability to combine them with basic knowledge of the world - formally known as "commonsense reasoning"
- remains an area of active research. Kawazoe has introduced the topic through a story, linked it to well known media topics like Watson, and briefly explained the current state of research in just sixteen pages.
And this is just chapter three!
If you're interested in AI and can read Japanese, I recommend you go and buy the book or check it out of your local library. If you can't read Japanese, well, I just hope you won't have to wait too long for an English edition. Ψ