Anyone reading Robert Heinlein’s The Moon is a Harsh Mistress remembers the lunar computer network that achieved sentience. This particular novel involved quite a bit of intrigue. Sentience and a quest for survival and independence are not disconnected phenomena. Like the 60's era Classic Rock, the teens of today have a high regard from some of the older speculative fiction (this particular work was originally published in 1966), and someone two generations removed from my formative years posed this question: “Will computers ever become sentient?”

I remember thinking about this a lot when I was going to SAC in the early 1970's, when I was riding a 70cc Honda and punching cards in the Nail Technical Center. One thing was clear was that there had to be a symbolic representation of real-world things, and that these real-world things would have multiple attributes. Object oriented programming, such as seen in C++, was still a few years in the future (the introductory C++ book was first printed in 1986). I knew what was needed, but not how to express it.

One of the little toys that emerged in the early days of the Macintosh was the first TV input device, and we had one of these where I worked at the time. It created all kinds of surreal effects, since it took several seconds to assimilate one frame, and most scenes had elements in motion. People’s faces and expressions were even more distorted than they are in so-called normal everyday life.

I realized from this that analyzing images of things in motion was computationally extreme (not only from multiple frames, but also from blur, shifting highlights, changes in scale, and partially obscured objects). Everyday things have their respective behaviors, so recognizing objects in motion simply requires cataloging everyday things, along with their static and dynamic properties. A leaf is green or brown, thin and light weight, and generally associated with trees. Lightweight, thin objects tend to get blown around by the wind and tend to follow tortuous paths in free-fall, so all these properties taken together could characterize a movie clip of leaves blowing in the wind.

Cataloging objects isn’t enough. It is helpful to predict interactions, and this is where I recognized the essence of sentience. People are continually thinking about how to improve their lot, find mates, or escape danger. In doing so, they play out scenarios involving multiple objects, characters, and circumstances. This “longhand” simulation is intended to identify possible outcomes, so that by the time a relatively rare event actually happens, the viewer is equipped to recognize the event for what it is. This is particularly useful if the event is someone named Bruno coming at you with a knife.

Since the number of real-world objects is large, and the number of interactions a power of all those objects, the human brain is left with the task of abstracting an infinitesimal amount of the universe. Interactions tend to get prioritized by their favorability or hazard; predicting interactions with lions and snakes gets higher priority than interactions with fireflies and okra. Recognition of edible fruits and nuts (and the context of their location and condition of ripeness or spoilage) is more important than recognition of rock types or astronomical phenomena. Recognition of and investments in women of child-bearing age or men who would be good providers is more important than people that don’t fall in those categories, except obviously for your existing children, parents, and people that beat you up when you’re trying too close to their mate.

A sentient computer, then, is “human” in that it has to concoct scenarios, place objects in interactive situations, run these situations to their conclusions, and catalog the outcomes. Prioritization needs to be given to things that confer survival or advantage, whatever that means to a dry, silicon based, electrically powered machine. The computer needs real world input devices to compare projected outcomes with observed outcomes. Where they agree, the model is reinforced; where they disagree, the modeling algorithm is defective and has to be refined. Getting progressively better algorithms sounds like learning in animals, or ‘New and Improved!’ in the commercial space.

Stories and storytelling, in this context, are a highly compressed shorthand for the real-world experience. Stories about lions told to children are better than having them watch (or experience) the real thing. A cursor scan of the grocery store bookshelf shows what the simulation priorities are: there are a lot more “romance novels” on bookshelves than there are “how to’s” on model railroads. Novels tend to contain a lot of violence, usually in the context of control of resources, whether these resources are one’s person, personal possessions, family, family assets, or tribal culture and tradition. There are a lot more Business Weeks then there are Newsweeks (i.e., more magazines dedicated to financial issues and stories than general news and politics) simply because a lot more people are concerned with making or keeping their money than they are with scandals in DC or starvation in Somalia.

All of this is surprisingly introspective. We “know” very little, if anything at all; this observation runs back to the Greeks and is a staple of any college humanities course.

Sentience is not the property of a computer so much as it is of a software application, and a network implementation is more likely and desirable. Heinlein had it right from the start.