MAIN IDEA:

The main idea of this book is to describe in very entertaining form various deficiencies of AI, as it is developed up until now and critic currently popular hype created by media around such systems. It is also designed to demonstrate complexity of such systems and difficult road ahead that needs to be travelled to overcome it.

DETAILS:

1: Mind the Gap

It starts with brief history of AI beginning in 1950s with its consistent over promising and underdeliver. The author provides a few examples of simple linguistic problems that easy for humans but very difficult for AI. He also provides a list of questions to ask in order to recognize overhype:

1. Stripping away the rhetoric, what did the AI system actually do here?
2. How general is the result? (E.g., does an alleged reading task measure all aspects of reading, or just a tiny slice of it?)
3. Is there a demo where I can try out my own examples? (Be very skeptical if there isn’t.)
4. If the researchers (or their press people) allege that an AI system is better than humans, then which humans, and how much better?
5. How far does succeeding at the particular task reported in the new research actually take us toward building genuine AI?
6. How robust is the system? Could it work just as well with other data sets, without massive amounts of retraining? (E.g., could a game-playing machine that mastered chess also play an action-adventure game like Zelda? Could a system for recognizing animals correctly identify a creature it had never seen before as an animal? Would a driverless car system that was trained during the day be able to drive at night, or in the snow, or if there was a detour sign not listed on its map?)

Finally, he discusses a number of areas where success is very close, but not achievable for a while, despite huge progress, such as driverless cars and makes the point that current AI with its trained neural networks becomes more functional and less understandable.

2: What’s at Stake

This chapter starts with the story of MSFT’s Tay – AI teenager who quickly learned lots of very bad staff from net and was shut down. It follows by discussion of AI’s lack of malice, personality and self-awareness. This makes them more controllable, but uncompetitive with humans in complex cognitive tasks. Author lists 9 specific risks linked to AI use. Author provides a couple of nice examples when non-human logic leading to logically consistent solutions unacceptable to people. Big part of it is AI substituting actual objectives of the task by some intermediate goal that is much easier to achieve. Nice example is soccer playing robot with set objective to touch ball as many times as possible, which start vibrating while touching ball. These problems are easily solvable, but practically non-predictable.

3: Deep Learning, and Beyond

Here author discusses massive move from classical – algorithmically programmed AI to deep learning self-programmed AI. The dramatic improvement in hardware power over the last 10 years greatly increased viability of this approach and shifted complexity to data selection. Author provided a nice graphic presentation of AI field:

Then he concentrates on Neural Networks, discussing their greed for data, opaqueness of results and fragility when it is not possible to understand how it achieved some weird conclusion. Here is representation of how it works:

4: If Computers Are So Smart, How Come They Can’t Read?

This is about another feature of AI – its inability of making sense out of reading texts or other forms of processing information. Author analyses a few examples of this happening. After this author discusses Google search algorithms and how they sometimes make mistakes inconceivable for humans, providing a few funny samples.

5: Where’s Rosie’?

This chapter is about progress in robot’s development or rather the slow tempo of such progress. So far we have Rumba, which is not that smart and hope for driverless cars that hit new hurdles all the time. Author discuses challenges of localization and situation awareness that robotics finds difficult to overcome and presents some “real life” scenarios to demonstrate impact.

6: Insights from the Human Mind

Here author reviewing some specific point, which make human intelligence so difficult to imitate:

• There are no silver bullets- complexity
• Excessive use of internal presentations
• Abstraction and Generalization
• Cognitive systems are highly structured
• Even simple aspects of cognitive systems require multiple tools
• Human thoughts and language are compositional
• Humans keep track of individual things and people
• Complex cognitive systems aren’t blank slates

7: Common Sense, and the Path to Deep Understanding

This chapters is about complexity of common sense and how difficult it is to recreate it via computers. Author reviews different attempts to recreate it algorithmically which generally fail because inapplicability of formal logic to common sense.

8: Trust

The final chapter discusses high requirements for AI system to be trustworthy and potential very high cost of errors even if probability of such errors is extremely low. Author discusses program verification methods, but admit that they are good only for simple systems like device drivers, but could not be used for complex AI, making the issue of the trust in such systems paramount to resolve before mass implementation.

Epilogue

At the end author expresses believe that eventually AI will become part of regular human environment and issues discussed in this book will be resolved, while presenting a bunch of new issues such as human employment that will need to be tackled.

MY TAKE ON IT:

I like multiple examples of clumsy AI provided in this book and generally agree that this technology is far from being close to full implementation as foreseen by Sci-fi authors and philosophers. However, I pretty sure that non-thinking, environment analyzing, and action directing systems such as required for driverless cars and based on Deep Learning are very close to implementation and will become trivial reality of everyday life. As to super sophisticated self-conscious system, I do not think they would go beyond some experimentation because in order to create such system, one would need to recreate complex experience similar to human life, which result in creation of just another human only on silicon instead of carbon base. I do not think that such artificial human would be superior in any shape and form to combination of regular human and computers with complex databases and AI analytical tools. Besides, similarly to what happened with nuclear weapons and conduct of war, not everything that could technically be done will be actually done due to multitude of ethical and common sensical limitation.