Tags

, , , , , ,

In chapter three of Turing’s Nightmares, entitled, “Thanks goodness the computer understands us!,” there are at least four major issues touched on. These are: 1) the value of autonomous robotic entities for improved intelligence, 2) the value of having multiple and diverse AI systems living somewhat different lives and interacting with each other for improving intelligence, 3) the apparent dilemma that if we make truly super-intelligent machines, we may no longer be able to follow their lines of thought, and 4) a truly super-intelligent system will have to rely to some extent on inferences from many real-life examples to induce principles of conduct and not simply rely on having everything specifically programmed. Let us examine these one by one.

There are many practical reasons that autonomous robots can be useful. In some practical applications such as vacuuming a floor, a minimal amount of intelligence is all that is needed to do the job. It would be wasteful and unnecessary to have such devices communicating information back to some central decision making computer and then receiving commands. In some cases, the latency of the communication itself would impair the efficiency. A “personal assistant” robot could learn the behavioral and voice patterns of a person more easily than if we were to develop speaker independent speech recognition and preferences. The list of practical advantages goes on, but what is presumed in this chapter is that there are theoretical advantages to having actual robotic systems that sense and act in the real world in terms of moving us closer to “The Singularity.” This theme is explored again, in somewhat more depth, in chapter 18.

I would not personally argue that having an entity that moves through space and perceives is necessary to having any intelligence, or for that matter, to having any consciousness. However, it seems quite natural to believe that the quality of intelligence and consciousness are influenced by what is possible for the entity to perceive and to do. As human beings, our consciousness is largely influenced by our social milieu. If a person is born or becomes paralyzed later in life, this does not necessarily greatly influence the quality of their intelligence or consciousness because the concepts of the social system in which they exist were founded historically by people that included people who were mobile and could perceive.

Imagine instead a race of beings who could not move through space or perceive any specific senses that we do. Instead, imagine that they were quite literally a Turing Machine. They might well be capable of executing a complex sequential program. And, given enough time, that program might produce some interesting results. But if it were conscious at all, the quality of its consciousness would be quite different from ours. Could such a machine ever become capable of programming a still more intelligent machine?

What we do know is that in the case of human beings and other vertebrates, the proper development of the visual system in the young, as well as the adaptation to changes (e.g., having glasses that displace or invert images) seems to depend on being “in control” although that control, at least for people, can be indirect. In one ingenious experiment (Held, R. and Hein, A., (1963) Movement produced stimulation in the development of visually guided behavior, Journal of Comparative and Physiological Psychology, 56 (5), 872-876), two kittens were connected on a pivoted gondola and one kitten was able to “walk” through a visual field while the other was passively moved through that visual field. The kitten who was able to walk developed normally while the other one did not. Similarly, simply “watching” TV passively will not do much to teach kids language (Kuhl PK. 2004. Early language acquisition: Cracking the speech code. Nature Neuroscience 5: 831-843; Kuhl PK, Tsao FM, and Liu HM. 2003. Foreign-language experience in infancy: effects of short-term exposure and social interaction on phonetic learning. Proc Natl Acad Sci U S A. 100(15):9096-101). Of course, none of that “proves” that robotics is necessary for “The Singularity,” but it is suggestive.

Would there be advantages to having several different robots programmed differently and living in somewhat different environments be able to communicate with each other in order to reach another level of intelligence? I don’t think we know. But diversity seems an advantage when it comes to genetic evolution and when it comes to people comprising teams. (Thomas, J. (2015). Chaos, Culture, Conflict and Creativity: Toward a Maturity Model for HCI4D. Invited keynote @ASEAN Symposium, Seoul, South Korea, April 19, 2015.)

The third issue raised in this scenario is a very real dilemma. If we “require” that we “keep tabs” on developing intelligence by making them (or it) report the “design rationale” for every improvement or design change on the path to “The Singularity”, we are going to slow down progress considerably. On the other hand, if we do not “keep tabs”, then very soon, we will have no real idea what they are up to! An analogy might be the first “proof” that you only need four colors to color any planar map. There were so many cases (nearly 2000) that this proof made no sense to most people. Even the algebraic topologists who do understand it take much longer to follow the reasoning than the computer does to produce it. (Although simpler proofs now exist, they all rely on computers and take a long time for humans to verify). So, even if we ultimately came to understand the design rationale for successive versions of hyper-intelligence, it would be way too late to do anything about it (to “pull the plug”). Of course, it isn’t just speed. As systems become more intelligent, they may well develop representational schemes that are both different and better (at least for them) than any that we have developed. This will also tend to make it impossible for people to “track” what they are doing in anything like real time.

Finally, as in the case of Jeopardy, the advances along the trajectory of “The Singularity” will require that the system “read” and infer rules and heuristics based on examples. What will such systems infer about our morality? They may, of course, run across many examples of people preaching the “Golden Rule.” But how does the “Golden Rule” play out in reality? Many, including me, believe it needs to be modified as “Do unto others as you would have them do to you if you were them and in their place.” Preferences differ as do abilities. I might well want someone at my ability level to play tennis against me by pushing me around the court to the best of their ability. But does this mean I should always do that to others? Maybe they have a heart condition. Or, maybe they are just not into exercise. The examples are endless. Famously, guys often imagine that they would like women to comment favorably on the guy’s physical appearance. Does that make it right for men to make such comments to women? Some people like their steaks rare. If I like my steak rare, does that mean I should prepare it that way for everyone else? The Golden Rule is just one example. Generally speaking, in order for a computer to operate in a way we would consider ethical, we would probably need it to see how people treat each other ethically in practice, not just “memorize” some rules. Unfortunately, the lessons of history that the singularity-bound computer would infer might not be very “ethical” after all. We humans often have a history of destroying other entire species when it is convenient, or sometimes, just for the hell of it. Why would we expect a super-intelligent computer system to treat us any differently?

Turing’s Nightmares

IMG_3071

Advertisements