Jim Baen's Universe

Nonfiction articles

Two Steps Toward General AI

Written by Stephen Euin Cobb

People are not comfortable spending time with someone who refuses to speak. Talkative people can sometimes become annoying, but silent people are just plain scary. Villains in the darkest horror movies often say nothing in scene after gory scene. Lack of communication is ominous. It’s a sign of danger. The concept is simple: my friends talk to me; my enemies don’t. This impression seems universal for all people, within all cultures, and may operate at a fundamental level within the brain, such that it can not be turned off, bypassed or lessened through training.

Those who want AI to become popular need to be aware of this. Because if it is true, the only way people will trust a future AI with a human level IQ is if it talks. And it’s going to have to talk a lot. People will trust it only if they feel they know what it’s thinking on a moment by moment basis. And the only way to be sure it’s begun talking enough to calm the nervous is when those who are not nervous begin telling it to shut up.

So let’s examine ways to approach the task of creating a talkative AI.

People often say, when it comes to learning language, the first goal of infants is to memorize words. While it is clear that infants are indeed memorizing words, this activity—which is observable—may actually be hiding the real goal—which is not so observable. The goal may instead be the same goal the infants will struggle with throughout life. And that goal is to predict.

For infants, whose minds are still simple, these predictions begin simply. So they first concentrate on learning to predict what words they are likely to hear next in a string of words spoken by those nearby.

Before they get good at this prediction skill, they will verbalize words and phrases in a parrot-like form of mimicry. This mimicry shows their enthusiasm to join in the creation of word stings but does not reveal any evidence that these young humans understand what their utterances mean.

Later, as they get good at this form of prediction, their utterances will contain a mixture of mimicry and meaningful content. At this point the humans are beginning to communicate and are generally considered to understand at least a little of what they are saying.

Once they become skilled at predicting what words will follow within a string of words they hear from others they will move on to a far more important area of prediction. The lifelong process of learning to predict the effect the various strings of words they speak will have on the people who hear them.

Responding well and appropriately to word strings from others is one half of the goal; saying the right thing for the current situation and the current set of listeners is the other. They will make many mistakes in their predictions. They will suffer mildly for some and greatly for others. But learning to predict the effect their word strings will have on their fellow human beings is fundamental to every young human's survival as a social animal.

And it will likewise be fundamental to the rise of AI.

Using this description of how humans learn language as a model, we can approach growing a General AI from scratch as the creation of a heuristic program designed first with the ability to predict what words are likely to follow in a string of words it hears from others; and then later with the ability to predict the effect the word strings it generates itself will have on those who hear them.

Since this second activity would be time consuming, it might be logical to have the heuristic watch a great deal of well-written movies and TV. “Well written” in this context means: accurate in its depiction of how people normally respond to word strings. And yes, this does sound annoyingly like the TV watching segment in the movie Short Circuit.

A much faster method, though not as good since it does not include body language or facial expressions, would be to have the heuristic program learn by reading novels. It might be possible for the heuristic to read a novel hundreds or even thousands of times faster than watching a movie and so develop this prediction skill far more rapidly.

But predicting word strings and their effect on listeners is not the only thing an AI or young human must learn. It must learn to plan ahead, to evaluate the risks of an action versus the rewards the action might yield, and to operate without hesitation even when it has vastly incomplete information which, for human beings, is very nearly all the time. For this article, however, we will examine only one more fundamental skill, and we will keep our focus on strings of words.

In conversation, humans use sentences of only a few types: 1) standardized greetings and responses. (“Hi, how you doing? Fine; how about you?”); 2) asking for something, or for someone to do something. (“Where's Frank? Can I borrow your pen? Let's go that way.”); 3) describing something we did or saw or heard or dreamt. (“I just bought a cool new phone.”)

The first type of sentence is just to get the conversation started and is so standardized and repetitive no one gives it a second thought. The second type is far more important and sometimes the most useful part of a conversation since it helps us get what we want, but it actually constitutes very little of a conversation. The bulk of conversations are composed of the third type.

The third type of sentence is usually thought of as description, and it is, but a lot can be revealed about how the brain generates description if we take a moment to look at it as translation.

Each time we describe the same thing to a different person, we customize the telling to suit them. If they understand the technical terms needed to grasp its full implications we use them; if they don’t, we use simpler words and might even leave out parts, especially if explaining those parts will take too long. For another listener we make it less gory or vulgar; for those with short attention spans we condense it; for those who love details we include them all. By rewording it, each telling is a translation designed to fit the audience that happens to be listening.

This is hardly news. We all do this, all the time, without even thinking about it. But the secret is not that we do this when we describe something to a second person. The secret is that we do this even when we describe it to the first. This is because, as we generate the first telling, we are translating into words the things we remember as reality.

For those of us who speak only one language, it may seem an odd notion that we’ve spent our entire lives translating. But if this is true—if conversation is mostly translation in disguise—this may mean we are closer to achieving General AI than many people presume. Translation software exists now. It’s clumsy, awkward and lacking in subtlety and nuance, but it does translate. On the other hand this software only translates one language into another. It does not translate reality into words.

Still, many of our conversations are not about the reality we’ve seen or experienced ourselves but about the things we’ve heard from others—friends, relatives, coworkers or even on the news. In those cases we are only translating words into other words.

It would seem that if today’s software can translate English into Chinese, it should be possible to make software which translates English into English only shorter, or with less jargon, or less vulgar, or whatever. Maybe, just maybe, the creation of such software might be a baby step in the general direction of General AI.

****

You can learn more about Stephen Euin Cobb here, or here.

And more about his podcast: The Future And You here, or here, or even here.