This book does not require any prior knowledge of neural networks, machine learning, artificial intelligence, or even computer science. Dr. Shane writes in a straightforward prose that is easily consumed - even by those unfamiliar with the math and science under the hood.
She begins with an explanation of Artificial Intelligence - what it is and why it is useful. She then covers some uses of AI, focusing on its limitations and misuses. Her samples include many unexpected results. The title of the book comes from an effort by a neural net to generate pickup lines after examining hundreds of actual lines.
Here are a few thoughts from Shane's book:
- An AI is very literal. It will try to solve the problem you give it - sometimes in unexpected ways. If you tell it to come up with a game-playing strategy that minimizes the number of times a player is killed, it may decide to hide in a corner and not move, which accomplishes the stated goal but is probably not an effective strategy for winning a game.
- An AI will take shortcuts if it can. In an experiment to identify the presence of sheep in a photograph, the AI noticed that nearly every photo of sheep also included grass. Since it was easier to identify grass than sheep, it concluded that any photograph of grass also included sheep.
- AI works best when it is given a narrow focus. It struggles if the problem is too broad. It is possible to create a bot that can have a conversation with a human, but that conversation will be far more meaningful if we train it to stick to a narrow topic. Try to train a bot to both take travel reservations and give relationship advice and it will likely fail at both.
- Because AIs are trained in a simulated environment, they may choose solutions that only work in that environment, but not in the real world. One experiment asked an AI to find the fastest way for a robot to get from one point to another. It concluded the optimal solution was for the robot to grow a long leg and fall toward the destination.
- Bias in input data can result in bias in predictive results. Train a system on existing resumes and hires and it may conclude that men are better hires than women because they were hired more often in the past.
As a result of these and other limitations, Shane concludes that we are unlikely to develop a general-purpose intelligence system, such as Star Trek's Data, 2001's HAL, or Terminator's Skynet any time soon. But that does not diminish the usefulness of the field, which can solve complex problems in imaginative ways. We just need to be aware of the pitfalls, so we can avoid them.