Browse Books

Reviewer: George R. Mayforth

Nick Bostrom has written an intelligent, thought-provoking book. I havent the space here for a detailed review, but hope to provide a flavor of what this book offers. Bostrom, with a wealth of detail, a bit of humor, and copious endnotes, touches on relevant topics regarding how human-made artificial intelligence (AI) could come into being and become far more intelligent than its creators, including: How and how quickly we might get there, whether by whole brain emulation (WBE) or creation of a human-engineered general artificial intelligence (GAI), both taking advantage of future advances in computing power; The many issues regarding control of such an entity, in particular, dealing with something that is far more mentally capable than a human being and making sure its efforts are beneficial to humanity versus possible malignant outcomes; and Strategies we might use to achieve a beneficial result. The primary problem in an effort such as this is that we must describe and discuss an entity that most likely would be completely alien. It would think faster and make analyses far more detailed than a human. Constrained as we are by human thought patterns, would we be able to comprehend, and thus describe, the thought processes of something vastly more intelligent than a human Would we be able to control it The components and a sense of the design of a superintelligent mind are aspects of discussing superintelligence. As Bostrom proceeds with his arguments, one can infer possible components of such a mind, but he presents no specific model. Having one would help in understanding why a particular attitude, such as animosity toward humans, might arise. For example, a thinking intelligence must have (at a minimum) sensory functions to gather data, a memory (database of information to store inputs and retrieve them), logic functions, comparison and pattern-matching functions, evaluation functions, goal-setting and decision-making functions, action functions, and the ability to extend its capabilities. Human beings (at least most human beings) also have a moral function that guides our thoughts of fairness, justice, right, and wrong. A WBE would presumably have all of this. Would we consider a GAI incomplete without a full complement of such components Could we imagine from this model that an AI, after reading human history and literature, would make a decision about whether humanity is good or bad, whether humanity should continue Bostrom repeatedly notes, quite rightly, that we should avoid anthropomorphism in discussing a superintelligent entity. He tries to do so, but not with complete success. He speaks, for instance, of a superintelligent agents goals, such as wanting to spread copies of itself through the universe using self-reproducing machines. This is, however, an expression of the fundamental goals of all biological life, survival and reproduction, that developed through eons of evolutionary processes. A WBE might possess such goals, but why would we necessarily expect a GAI to have them unless so endowed by humans He also speaks of an AI that might consume the worlds resources to meet its own goals, possibly eliminating humanity, either willfully or as a side effect. A WBE would have a human moral component that might mitigate this problem. On the other hand, if we could give a GAI a moral component, this might alleviate the problem. Without it, the GAI might be unconcerned that allocating resources for its own use damages humanity. There are moral issues. Would an advanced AI be morally or legally equivalent to a conscious human What would it mean to power off or destroy such an AI Could mental anguish be inflicted on such an entity If such AI entities are instantiated and decommissioned as workers for a business, what are the moral and legal implications Such issues should be decided before an AI of this capability is created. There are political issues. Would it be better to have a single AI system versus multiple different competing systems Does it make sense to have international collaboration on the development of an advanced AI Given our experience with nuclear technology, it is difficult to imagine all actors on the world stage agreeing to be restricted by such a collaboration. This is an important book that raises many issues that must be addressed before the creation of an advanced intelligent entity. One could argue that Bostrom takes caution to excess, but perhaps given the possible negative scenarios, he is correct to do so. We face, after all, a technology that is unknown, possibly unfathomable, and possibly dangerously uncontrollable once it is active. More reviews about this item: Amazon, B&N, Goodreads, NPR Online Computing Reviews Service